Changelog

A changelog you can read from the app, and a way to reach the founder.

Added

• Changelog in the app. What's new now has an "Earlier releases" link, and About has a Changelog link, that show every release newest first, fetched from anjadhe.ai/changelog only when you open it. Offline, the shorter list the app carries with it is shown instead.
• Contact the founder. A direct email link in About, in Settings › Send Feedback, and in the What's new and Changelog views. Bug reports and "I'm lost" messages go straight to the person who wrote the app.

Changed

• The website's download page can now email you release notes: one email per release, no account, unsubscribe deletes your address. The list lives on Anjadhe Connect, whose source stays public.

Notes editor fixes and a close-window guard for running AI work.

Added

• Markdown pastes as formatting. Text copied from a .md file lands in a note as real headings, lists, quotes, code blocks, rules and bold/italic/links, not as source with # and ** in it.
• Normal text in the note editor: a T button on the selection toolbar, a "Text" item in the / menu, and Cmd+Shift+0 turn any heading, quote or code block back into a plain paragraph and strip inline formatting.

Changed

• Closing the app asks when AI work is running. Closing the window, Cmd+Q and Force Reload now show the same kind of warning Cmd+R already did, naming what would stop (an assistant reply, a task run, a build). Stay to keep working, Leave to go anyway. Ambient email analysis, which resumes on its own, still never prompts.

Fixed

• Pasting several lines of plain text into a heading no longer turns the whole paste into one giant heading; each line becomes its own paragraph.

Cloud privacy: when the model runs off your Mac, you can see exactly what left, decide what background work may send, and check the server.

Added

• "Left this Mac" ledger. Settings › AI Assistant › LLM Logs filters to every request that went to Anjadhe Cloud, your own API key or your server, with its destination and the exact messages it carried. AI Activity rows carry a destination chip, and the composer's model chip shows a small arrow whenever answers come from off this Mac.
• Cloud Privacy switches (Settings › AI Assistant › Cloud Privacy): one switch per kind of data that background work — email insights, thread status, routines, reviews — may send to a cloud model. Journal and Wellness are off by default. Held-back work is skipped and shown in AI Activity as "kept on this Mac", never dropped silently. Chat is never gated. A one-time note explains this the first time you pick a model that runs elsewhere.
• Liabilities in Portfolio. Mortgages, loans and credit lines, optionally secured by a property; net worth headlines the masthead and home card when debt exists. Sync-safe like every other portfolio record.
• A cloud-privacy help topic for the assistant (and on anjadhe.ai/help).

Changed

• Less leaves. Email bodies lose quoted history, signatures and tracking links before background analysis goes to a cloud model, and background prompts on a cloud model are capped in size.
• Anjadhe Cloud card states its commitments in checkable sentences, names DeepInfra as the inference provider under a no-retention agreement, and shows the commit the server is running (Connect v1.5.0's /v1/version) for comparison with the public repository.
• Models are grouped by where they run — this Mac, your server, API providers, Anjadhe Cloud — in Settings › Models and in the model dropdown alike.

First-run welcome pages across the apps, and task runs that pace themselves on Anjadhe Cloud instead of tripping its rate limits.

Added

• Apps greet you now. An app you haven't set up yet opens on a short welcome — what it does, what connecting gets you, and the door in — instead of an empty pane. Email, Calendar, Journal, News, Notes, Portfolio, Reader, and Writing Voices all have one.
• Quick-start suggestions on Home. A few prompt pills under the composer suggest things worth asking; the set stays put all day, so it can be used by muscle memory.

Changed

• The app's website links (Terms, Privacy Policy, help) point at anjadhe.ai, the public domain.

Fixed

• Task runs no longer rate-limit themselves on Anjadhe Cloud. The task engine paces its calls on a metered brain and, when the service asks it to slow down, waits exactly as long as asked and retries — no more runs dying mid-step over finished work.
• Truthful task reports. When a run consolidates per-item work into one deliverable (say, "write the document" planned as per-item edits), the items now settle as "unverified" instead of the whole step reading failed over completed work.
• No more "No models installed" banner over a working cloud model. The local-engine setup nudges no longer apply to remote engines.
• The assistant acts instead of narrating. "Save that" or "reschedule the plan" now performs the action rather than announcing it and stopping.

Email analysis now paces itself on Anjadhe Cloud so a fresh account connect no longer trips the service's rate limit, and the app launcher gets a cleaner "More apps" page with a leaner default set of apps.

Added

• More apps page. The launcher's on/off surface is now one card per app — icon, name, a one-line description, and a switch. Reached from the sidebar's "More apps" row, the app switcher, and Settings › Appearance. Turning an app off only hides it from the launcher: its data stays, and Cmd+K still finds it by name.
• A leaner default. Wellness, Web Browser, AI Activity, Help, and About start switched off on new installs; one click on the More apps page brings any of them back — and the choice sticks.

Changed

• Bookmarks live inside the Web Browser. Every new tab shows your bookmarks, and its "Manage" button opens the full Bookmarks page.
• The Actions app reads "Tasks" in breadcrumbs and help — the same app, named for what it holds.
• The home feed's welcome post now explains routines — what they are, the ready-made ones, and how to create your own by asking the assistant.

Fixed

• Connecting an email account no longer trips Anjadhe Cloud's rate limit. Insight analysis spaces its requests out on a cloud model, and when the service does ask it to slow down, every part of Email AI — insights, reservation details, thread judgments, inbox bundles — waits and retries instead of skipping mail. Before this, a burst of 429s could park thread judgment and bundling for the whole session and silently cost a booking its reservation facts.
• Tasks list dates line up. A long date label ("Annually (Dec 13)") widened its own row's date column, pushing that row's title out of line with the rest.
• Blank-slate test instances no longer see real documents. ANJADHE_DATA_ROOT now also redirects the ~/Anjadhe/ content tree (library, user apps, artifacts).

Your Mac's Apple apps join in: import Reminders into Tasks, Notes into Notes, and see this Mac's Apple Calendar events in the Calendar — all read locally from the Mac itself, no Apple ID or password involved. Tasks also gain a tag system, born from Reminders lists.

Added

• Apple Reminders → Tasks. Settings › Accounts › Apple apps imports this Mac's open reminders as tasks — due dates, times, and repeats (daily, weekly, monthly, yearly) included. A list picker chooses which Reminders lists come along, each task is tagged with its list name, and completing a reminder in Apple's app marks the task done here on the next import. Read-only: nothing is ever written back.
• Apple Notes → Notes. Notes import with their text, tagged apple-notes; locked notes are skipped. Meant for one Mac — imported notes reach your other Macs through Anjadhe's own sync.
• Apple Calendar in the Calendar. Events from this Mac's iCloud and local calendars appear as a read-only mirror with their own "Apple Calendar · this Mac" row in the accounts rail, scoping like any account. Google-account calendars on the Mac are deliberately excluded — Anjadhe syncs Google Calendar itself.
• Task tags. Tasks carry tags now: a Tags row in the task editor (search existing tags or create one in place, the notes editor's picker), quiet tag chips on task rows, and a Tags section in the Actions nav — "All tags" plus a row per tag — for finding tasks by tag.

Changed

• Turning an Apple import off removes what it brought in. The data still lives in the Apple app, and switching back on re-imports it — the same model as Gmail, where the source owns the mail.

Fixed

• Notes: a tag view with the sidebar hidden had no way back. The list now shows a "← All notes" chip whenever a filter is active and the filters sidebar isn't visible — including the empty-tag case, which was a dead end.
• Memory: the assistant's background note-filing no longer stalls on a big backlog. The compaction pass sized its reply for a small batch, so a backlog truncated the output and nothing was ever absorbed; the reply cap now scales with the batch and partial results are kept instead of discarded.

Anjadhe Cloud grows a model choice: the service can offer several hosted models, and you pick which ones to add.

Added

• Choose your Anjadhe Cloud model. Add model › Anjadhe Cloud now lists the models the service offers — each with a name and a one-line description — instead of a single fixed entry. Every pick becomes its own entry in Your Models, so cloud models sit side by side with local ones and switch from the model chip like anything else. The lineup comes from the server's own catalog, so new models can be offered (or swapped for better weights) without an app update; if the catalog is unreachable the step falls back to the original single model.

Changed

• Cloud model names follow the server. If a hosted model is renamed in the service catalog, existing installs pick up the new name at next launch — on the model chip, the Settings cards, and feed post bylines.
• The Anjadhe Cloud Manage panel names the server it talks to — normally api.anjadhe.com, and explicitly marked when a test run overrides it.

A small release: routines ride out AI outages instead of skipping a day, and an Email AI insight shows the message it came from.

Changed

• Routines retry through an AI outage. A scheduled digest whose run fails because the model is unreachable used to post an error card and give up until the next scheduled day. The card still posts, but the run now stays queued on a backoff ladder (15 min, 30 min, then hourly, about six hours of cover) until it succeeds or a newer scheduled run supersedes it. Configuring a model immediately retries whatever is waiting, and a successful post folds the error card away.
• Email AI shows the source message. Opening an insight now renders the full original email below the insight detail, with the same sanitized treatment as the Inbox viewer, so the summary can be checked against the mail itself.

Fixed

• Disk space is reported correctly. macOS df output for the sealed system volume shows the OS-only Used/Capacity columns (~3% used on a nearly full disk), and the assistant could invert that into "97% available". The true free percentage is now computed from the real numbers and attached to the result.

Trips assemble themselves from your reservations, records join the conversation with @, and a whole goal plan can be rescheduled in one step.

Added

• Trips. Reservations that belong together — the flight out, the hotel, the rental car, the return leg — are grouped into a trip by their dates; nothing to set up. A coming trip gets its own card on Home, and the Trips entry in Email AI's sidebar opens the index: upcoming trips soonest first, past trips below. Opening a trip shows every reservation in one place, including bookings made months ahead whose email has long since aged out of the 90-day insight window.
• Mention a record with @. Type @ in any assistant message box to pull up your tasks, goals, notes, routines, strategies and accounts; the first word can narrow the type ("@goal marathon"). Picking one attaches the conversation to that record — current state and its decisions included — with a banner naming the attachment and an ✕ to detach.
• Reschedule a whole plan at once. "Start this goal today" now shifts every task in the goal's plan in a single step — spacing preserved, weekday cadence optionally kept — with the full before/after preview in the confirmation dialog. No more one-task- at-a-time edits.
• Paste images into notes. An image or screenshot pastes straight into the note body, is stored inside the note (scaled down to keep notes light), and syncs with it.
• Skipped insights are inspectable. Everything Email AI learned to skip is listed under Email Settings › Skipped insights, each with a "Show again" undo; right-click a message in the inbox and choose "Analyze with AI again" to re-run it after unlearning.
• Study depth. Settings › Reader & Writing Voices sets how much of a voice's documents one Study pass reads.

Changed

• Reader search runs on a real vector index — semantic search over a large library is much faster, with the same results.
• Journal panes collapse like Notes' now, and both Notes and Journal remember your reading-pane preference per Mac.
• Email AI rows read like the Inbox's: the whole row responds to hover and carries the same hover "Mark read" pill.
• Goals' left nav drops completed goals, so the group list is what is still in play.

Fixed

• HTML-only email is finally read in full. Messages with no plain- text part (a large share of commercial mail) were being analyzed from their one-line preview snippet, which is why some obvious insights came out wrong or thin. The full message text is now derived from the HTML for analysis, matching, and chat over the open message. A purchased booking also files as a reservation now, not a receipt, so it joins its trip.
• Long emails and long notes no longer get cut off mid-read when the assistant works with them — it pages through the rest instead.

The Reader arrives and Writing Voices turns on for everyone, Wellness grows from a health log into a coach, and Notes gets collapsible panes.

Added

• Reader — a new app for your imported documents (reports, essays, papers, statements). Import files, drag-drop, or drop them into the folder in Finder; search everything by meaning, open a document in place with matches marked, and Ask about this document starts a chat that stays attached to it — the assistant answers from the document's actual text, never from memory of a familiar title.
• Writing Voices is on for everyone — no more experimental flag. Two sample voices are bundled (Mark Twain and Abraham Lincoln, from their public-domain works) and import with one click, so you can hear what a studied voice sounds like before teaching your own.
• Routines can speak in a voice. A digest routine — a daily review, a morning motivator — can carry a Writing Voice for its posts. Pick it on the routine's form; the assistant's own voice stays the default.
• Wellness is a coach now. The page opens with a quick-log line — type "BP 128/82 after my walk" or "strength training 45 min" and the assistant logs it with every detail kept — over a Today strip and an attention briefing (streaks, trends, what you usually log and haven't) computed from your entries, never guessed. One click arms a weekly honest review or a daily motivator, optionally in a voice you taught. Habit-level only; it never diagnoses.
• The assistant can correct the log. Fix a mistyped reading or delete an entry by asking — corrections update the entry in place, and deletes ask first.
• The assistant can delete notes (with a confirmation, like every delete).
• Search in the AI logs. Settings › the model and search logs now filter as you type — who called, which model, what was sent, what came back.

Changed

• Writing Voices slims to its name: the voices list and each voice's documents. Global search, other documents, and the corpus reader moved to Reader — the two apps are lenses over one folder, and every document is visible in both.
• Notes panes collapse individually. The filters column and the note list each get their own toggle, remembered per Mac; opening a note from elsewhere (a goal, a record link) folds them just for that visit.

Fixed

• Clicking between paragraphs in a note no longer adds empty lines. The editor only inserts a paragraph where the caret truly had nowhere to land (between two code blocks, images, or tables).
• "Summarize this document" over an open document now works even when phrased loosely — the chat carries the document itself, so the assistant no longer answers "I don't have a tool for that" or summarizes the famous book instead of your copy.

Fixes a launch crash in v0.1.0-alpha.33, which was pulled minutes after publishing. Everything alpha.33 introduced ships here instead.

Fixed

• The app launches again. alpha.33 shipped without a new module the app requires at startup (llamacpp/embed-manager.js missed the packaging list), so it crashed before drawing a window — which also meant the auto-updater couldn't run. If you updated to alpha.33, download this version manually; from alpha.34 on, the release pipeline verifies every startup module is packaged (and launches the built app) before a release counts as published.

Added

• Writing Voices (experimental, off by default — Settings › Developer › Experimental features). Create a named voice — My voice, or anyone whose writing you have — give it documents (posts, essays, books, speeches), and press Study: the assistant writes a short, editable style guide plus real passages it will imitate. Your edits to the guide survive re-studies; pin the passages that carry the voice best. Draft in this voice… opens a chat already primed to write that way, and the draft names the documents it drew on. Everything is indexed on your Mac by a small local model; semantic search over your documents is a one-time ~330 MB download in Settings › Writing Voices. Searching opens the matching document right there — full text, terms marked — with doors to the original file, and deleting a document moves the file to the Trash, never a hard delete.
• AI Assistant in the sidebar and app switcher. It sits alone at the top, right under Home. If you had hidden the old tile, it comes back once — hide it again and that sticks.
• Email AI: a Connect an account button on the empty state, so getting started doesn't require finding Settings first.

Changed

• Connecting a Google account starts the first email and calendar sync immediately — from Settings as well as setup, and when a service is toggled back on. No more empty app until you open it.
• Email AI overview rows open the insight in place instead of jumping into its folder — the nav stays put, and back returns to the overview.

Writing Voices arrives as an experimental app, the AI Assistant takes its place at the top of the sidebar, and connecting an account starts syncing right away.

Added

• Writing Voices (experimental, off by default — Settings › Developer › Experimental features). Create a named voice — My voice, or anyone whose writing you have — give it documents (posts, essays, books, speeches), and press Study: the assistant writes a short, editable style guide plus real passages it will imitate. Your edits to the guide survive re-studies; pin the passages that carry the voice best. Draft in this voice… opens a chat already primed to write that way, and the draft names the documents it drew on. Everything is indexed on your Mac by a small local model; semantic search over your documents is a one-time ~330 MB download in Settings › Writing Voices. Searching opens the matching document right there — full text, terms marked — with doors to the original file, and deleting a document moves the file to the Trash, never a hard delete.
• AI Assistant in the sidebar and app switcher. It sits alone at the top, right under Home. If you had hidden the old tile, it comes back once — hide it again and that sticks.
• Email AI: a Connect an account button on the empty state, so getting started doesn't require finding Settings first.

Changed

• Connecting a Google account starts the first email and calendar sync immediately — from Settings as well as setup, and when a service is toggled back on. No more empty app until you open it.
• Email AI overview rows open the insight in place instead of jumping into its folder — the nav stays put, and back returns to the overview.

Every app is now one click away, all the time: a sidebar down the left of the window.

Added

• A sidebar listing every app. It runs down the left of the window, grouped by what each app is for, and marks the one you are in. Home is its first row. It shows exactly what the titlebar's app grid shows — hide an app in Customize and it leaves both, while staying findable by name in Cmd+K.
• Collapse it when you want the room. The Collapse control at the foot of the sidebar narrows it to icons and keeps it that way until you expand it again. The choice is per-Mac, like the theme. On a narrow window it collapses on its own, and on a very narrow one it steps out of the way entirely — Cmd+K and the titlebar grid still reach everything.

Changed

• The help the assistant answers from covers getting around. Asking it how to move between apps now describes the sidebar, the wordmark as the way home, and the back button, instead of only Cmd+K.

New email reaches Anjadhe within a minute, and chats pinned to a task or goal stop losing track of their subject.

Changed

• New mail shows up within a minute. Anjadhe used to slow its Gmail checks to every 3–10 minutes while you were away from the Mac — which is exactly when a confirmation email lands in Gmail and sits invisible here. It now checks every minute for as long as the app is running (each check is one tiny "anything new?" call), and checks immediately when the Mac wakes from sleep.
• The About view carries the sharpened pitch. The privacy promise section gains the breach/subpoena line and the unplug-the-network test, and the footer tagline becomes the ownership line. The footer also links the Terms of Service and Privacy Policy — the app never linked its own terms anywhere before.
• The Strategy page says what it is not. A quiet line at the foot of the strategy list and detail states Anjadhe is not an investment adviser, matching the new investment-tools clause in the terms.

Fixed

• Chats attached to a record keep their subject. A conversation pinned to a task, goal, note, routine, strategy or account answered "which task are you referring to?" when continued from the Assistant view. It now rebuilds the record's live context wherever the chat is continued. Chats over untrusted content (incoming mail, web pages) deliberately do not.
• A failed mail check can no longer silently stop mail checking for the rest of the session — the next check is always scheduled, even when one fails.
• Release pipeline: the asset-verify pass now runs and repairs a partial publish regardless of the build tool's exit code (alpha.30's own publish needed a manual rerun for exactly this).

Decisions: the plans you settle with the assistant stop evaporating into chat history — plus a speed fix for Goals and sturdier links in briefings.

Added

• Per-record decisions. A decision is a dated instruction pinned to one specific record: a strategy's "invest $2,000 monthly through October, split 60/40", a task's "never move this past Friday". Settle one in chat and the assistant asks before saving it (you'll see a quiet "Noted on …" confirmation); from then on it re-reads that decision every time it touches the record, in any conversation, so a plan made in one chat governs the next. Every record that can carry decisions shows them on its page — the task editor, goal details, strategy details, a portfolio account, a routine — where you can also add or delete them by hand. Decisions sync between your Macs. Chats that read untrusted content (web pages, incoming email) can neither read nor write them.

Fixed

• The Goals page is fast again. Opening a goal with a long task list re-read the whole schedule from storage once per task — and the page's nav and cards repeated the work per goal. It now reads once.
• Task links in briefings survive a mistyped link. A routine's briefing sometimes cited tasks whose links rendered as dead underlined text because the model wrote the link URL slightly wrong. The common misspellings are now understood, and a link that still can't be understood keeps its original form recorded so it can be diagnosed.
• The titlebar Back button's arrow icon now actually renders alongside the name of where you came from (it silently fell back to the house icon before).
• The Morning News routine can reach your News feed again — its digest is grounded in the headlines you follow instead of a plain web search.

A small release that opens the strategy record up and makes the app a better citizen of Anjadhe Cloud's rate limits — plus a privacy pass over the app's own console logs.

Added

• Every strategy has its own page. Clicking a plan's name on the Strategy page opens the whole record: the computed verdict, every saved property, the target mix and guardrails, the accounts actually following the plan (each one a link to that account), and the plan's change history — recorded since the first save, shown for the first time. The question pills underneath fit the plan's state: a draft offers to be finished, a drifting plan offers the way back, a plan nobody follows offers to be assigned.

Changed

• Anjadhe Cloud rate limits are paced, not fought. When the cloud says "slow down", background email analysis now pauses exactly as long as asked and resumes where it left off, instead of burning its retries into a closed window. Scheduled routines treat a rate limit as "try again shortly" rather than posting an error card to your feed.
• The app's own logs no longer quote your content. Console output (visible when Anjadhe is launched from a terminal, and capturable by anything that redirects it) used to include snippets of prompts, model replies, email subjects, and resolved article URLs. Those lines now log lengths and shapes only; the full traces still exist for debugging but are off unless explicitly enabled.
• A brand-new install's first briefing welcomes you instead of reporting that nothing is due, nothing is overdue, and there are no goals — it says what will land there each morning and suggests a first step.

Fixed

• Linking to a note whose title was empty (possible for notes created on the phone) no longer breaks the link picker's search.

The release where the AI story got honest about small Macs. Local models below 32 GB of memory never worked well enough to stand behind, so they are no longer offered — and the new Anjadhe Cloud is the suggested brain for those machines instead. The first-run wizard was rewritten around that story, in plain language a person would actually say. Alongside it: recurring events landed in Calendar, and every detail page grew a door to the assistant.

Added

• Anjadhe Cloud — an open-weight model hosted by Anjadhe, for Macs that can't run a capable model locally: no account, no key, a free monthly allowance of 1,000 AI requests while it's in preview. What runs on it goes to Anjadhe's server, then to an inference provider under a no-retention agreement, without your identity — nothing is stored, nothing is used to train models, and the service's source code is public so that can be checked. Never a default or a fallback: it exists only if you add it.
• Recurring events in Calendar. The event form has a Repeat row (daily, weekdays, weekly, monthly, yearly), editing an occurrence asks whether the change is for that event or the whole series, and the event detail says how the series repeats. Deleting already asked; editing now does too.
• "Ask about this…" everywhere. Task, goal, calendar-event, open-email, and email-insight details all carry a pill that opens the assistant with that record as context and the composer empty — you ask your own question, nothing is pre-written for you.
• Executed trades in email become portfolio transactions. A brokerage confirmation's trades are captured with the insight and carry an "Add to portfolio" button — the same review dialog as ever; nothing writes itself.
• Task rows multi-select. The row checkbox selects; a bulk bar deletes with one confirm and one undo. Completing moved into the row menu. The assistant can also delete a goal together with its tasks, behind the same ask-first permission as every delete.
• A little confetti when you complete a task.

Changed

• Local models are no longer offered below 32 GB of memory. The small models that fit 8–16 GB Macs didn't work well enough, and offering a model that disappoints is worse than offering none. Macs under 32 GB are offered Anjadhe Cloud first — with a server you own or your own OpenAI/Anthropic key as the other homes — in first-run setup and in Settings alike. Existing installs keep any small model they already added.
• The first-run wizard now explains Anjadhe the way a person would. The welcome page introduces a blank canvas with an assistant you can talk to and three concrete things to ask it; the model step opens with "a large language model is your assistant's brain, and where it runs is where your data goes"; the storage step says plainly that there is no remote database, that your data isn't our product, and that the source code is public.
• Add a model is a full page (was a modal): the homes for the model on the left — each naming where your data goes — and the chosen step beside them.
• The model list says "Local" for models on this Mac, instead of naming the engine.

Fixed

• A stranger who learns a Mac's routing id can no longer evict it from the mobile-sync relay. A connected Mac's slot is now only replaced by a reconnect proving it holds the same key; knowing the id alone was enough before.
• Anjadhe Cloud replies report the product model name, not whichever open-weight model the service happens to run upstream.

A short release, and all of it is about the promise the app makes rather than what it can do. Preparing the Google OAuth verification material meant reading our own claims against our own code, which turned up three places where the two had drifted: a launch quietly called Google for fonts, every backup put the whole database into iCloud unencrypted for a moment before encrypting it, and two model downloads shipped without a checksum to verify them against.

Fixed

• Backups never put the plaintext database in iCloud, even for a moment. The backup ran VACUUM INTO straight at the iCloud folder and encrypted the file afterwards, in place — so on every backup, hourly by default, the whole unencrypted database sat in the synced folder for the length of the encrypt. Whether iCloud uploaded that version before it was overwritten was a matter of timing nobody controls. Worse, the file stayed in cleartext if anything failed in that window, which the fail-closed encryption path made a real possibility rather than a theoretical one. Backups now build in a staging folder outside iCloud, are encrypted there, and only the finished ciphertext moves across; a backup that cannot be encrypted is not written at all.
• The two most common models are now checksum-verified. gemma4:12b-it-qat and gemma4:e2b-it-qat shipped with an empty hash pin, so their downloads were accepted without being checked — the two models a new install is most likely to pick were the two going unverified. All nine models with weights are now pinned.

Changed

• Starting the app no longer contacts Google for fonts. The interface typefaces were fetched from Google's font service on every launch, which meant an unconditional request to Google, carrying your IP address, before you had connected any account at all. The fonts now ship inside the app. The same request was being made again on every user-built app you opened.
• The privacy policy now names every service the app can talk to, what it sends there, and why — replacing a claim that only three destinations existed, which was not true. It also states plainly that your data is never sold, and it now lists the exact OAuth scopes the app requests rather than an approximation of them.

Two days of work on the mail side of Anjadhe, and it shows. Email AI is now an app of its own with a front page that says what needs you. The Inbox learned which threads need a reply and which are waiting on someone else. Routines and Automations became one thing, with an engine that fires when mail arrives instead of on the next poll. And News grew a search box, a bigger topic catalog, and an assistant that reads the article before it answers.

Added

• Email AI is its own app. What was a tab is now a full app (reachable from the app switcher and Cmd+K): insight folders on the left, that folder's rows in the middle, the open insight on the right. It opens on an Overview that leads with what is coming up (soonest first) and what arrived since you last looked, each row a door straight to its insight. The reading pane is a preference here too, folders never disappear just because they are empty, and the unread filter is the same envelope button the Inbox has.
• Needs reply / Waiting on them. Two computed lists at the top of the Inbox sidebar: mail a person sent you that you have not answered, and mail you sent that has been quiet for three days. Who wrote last is arithmetic; whether it is really a person expecting an answer is judged by your AI. Reply and it clears itself; archiving clears it too, and a right-click offers "No reply needed" / "Stop waiting on this" for the rest.
• Right-click works on mail. Inbox rows have a context menu: mark read/unread, archive, move to a bundle, and the thread-state exits above.
• The AI reads attachments. The assistant can open an attached PDF (scans included), spreadsheet, or document, and the insight engine can fill a bill's missing due date or amount from the attachment the mail points at. It fills blanks only; an attachment can never create an insight or make something actionable.
• Routines that act, triggered by things. Automations merged into Routines: one page for everything the app does on its own. A routine can run on a schedule or when an email or file arrives, as a read-only digest or a full task run that can act. Email triggers fire as mail lands, once per message; a routine's runs are listed on its own page with a per-run log; and a Test trigger button shows what would fire without running anything. Setting one up is a conversation: "Help me set up a routine" walks you through it.
• Search in News. A search box in the News header filters the current headlines (the last 48 hours of every topic you follow) and your Saved stories. Topics also carry 20 stories each now instead of 10.
• A topic catalog you can search. The topics dialog groups its suggestions (World, Tech & Science, Business, Sports, Culture & Life) with a filter box, and anything you type is a topic too. Up to 15 topics, and the limit is enforced out loud instead of silently dropping the ones past it.
• Send Feedback. Settings has a Send Feedback card that reaches the maker of Anjadhe directly, also available from the app menu and Cmd+K. It shows exactly what a send carries: your message, the app version and platform, and an app-details line you can untick. Attaching your analytics id so a report can be matched with usage data is a separate checkbox, off unless you check it, shown only if you opted into analytics.
• A ticker's page links out. A portfolio symbol's detail page opens the same symbol on the sites that trade or chart it.

Changed

• Insight folders say one thing each. Bills, Receipts, Renewals, Appointments, Reservations, Deliveries, Deadlines, Codes, Security — one noun per folder, each stating its own definition, so where a mail lands is predictable. Marketing is not an insight, however much it talks about renewals or discounts. Existing insights were refiled automatically.
• Your AI reads every incoming email now. A keyword pre-filter used to decide which mail deserved analysis, and it missed real bills and receipts over phrasing. The model itself is the judge now. (Installs using a paid API key keep the thrifty pre-filter, because there each call is your money.)
• An insight that became a task stays visible. Email AI shows every insight from the last 90 days, including ones that turned into tasks; the task and the insight link to each other.
• Opening a message counts as reading it sooner. In the reading pane after one second instead of one and a half; opened full-width, instantly.
• "Ask about this" in News reads before it answers. The chat starts with the article's own text (not just the summary), and when web search is on, the assistant searches for background and newer developments instead of trusting what it remembers.
• One market review. The weekday pre-market, midday and close reviews collapsed into a single Market Review timed to your market's close, expressed in your own timezone.
• Cmd+R asks before it kills AI work. Reload used to silently end whatever the assistant, a routine, or a re-analysis run was doing. Now it says what is in flight and asks; work that can resume, resumes from any page.
• A widget's color means the data's state (amber warns, green is good, red alerts), never which app it belongs to.
• Feedback tells the truth about your Mac. Every Mac used to report itself as Intel; reports now carry the real chip and memory.
• Choosing a remote brain frees the local model's RAM instead of keeping the local engine loaded behind it.

Fixed

• A custom news topic added past the fetch limit silently never loaded; every followed topic is fetched now.
• Websites that use bot protection (Zillow and many others) load in Browse again; sites that hard-block from the network get an honest banner with a way to open them in your default browser.
• An insight you delete stays deleted instead of returning on the next analysis pass.
• Following an insight into your mail is no longer a one-way door — Back returns to the page you came from, and the message no longer flashes and vanishes on the way in.
• Verification codes no longer become scheduled tasks that fire after the code is dead; they file under Codes with the code up front.
• API keys saved in Settings no longer read as corrupt after a relaunch (a startup-order bug in how the keychain was opened).
• An open Anjadhe window keeps the Mac from idle-sleeping mid-run, so long tasks finish.
• Cmd+K opens the palette even while writing a note; linking inside the editor moved to Cmd+Shift+K.
• A task due today can hop to tomorrow from its row menu.
• A routine digest that fails on a flaky connection retries quietly instead of posting the error to your feed, and a run interrupted by a restart picks itself back up.

This release is about the planning side of Anjadhe getting simpler and more honest. Focus areas are gone, replaced by a plain group label on each goal. Your portfolio plan became a page you can actually read, with the numbers computed rather than narrated. Scheduled prompts are now called routines and have a proper table. And the assistant got substantially better at finishing multi-step work instead of claiming it did.

Added

• A Plan page for your portfolio. The strategy used to live only in conversation. Now Plan in the accounts nav shows it: the plan in your own words, a target mix table with each sleeve's target against actual, a bar marking the tolerance band, and what to trim or add in dollars to get back on target. Under it, your guardrails with a pass or fail beside each, and a one-line verdict: On plan, Drifting, or Off plan. Every number is computed by the app with the same arithmetic the assistant uses, so the page and the chat cannot disagree. Nothing on the page is a form: anything that changes the plan is a button that hands the question to the assistant, and Ask about this strategy opens a chat with that plan already in context.
• Weekday reviews of a plan, beside the plan. "Review it every weekday morning" schedules a routine, and its latest read is annotated in the margin next to the strategy it is about. Goals work the same way with a weekly review.
• Goals can be deleted, and so can strategies. A plan you no longer follow can be removed from its card, and the assistant can do it too.
• Quick-add understands you. Type a task in plain language in Actions and the assistant can turn it into a proper task with a date, a repeat, and a goal.
• An engine update when there is one. When a newer llama.cpp build is available, Settings, AI Assistant, Your Models shows an Update engine row. It installs the same way the first one did.
• Search finds your portfolio accounts, both in Cmd+K and for the assistant.

Changed

• Focus areas are gone; goals carry a group label. There is no longer a set of focus areas to create, colour and maintain. Each goal has a plain Group field ("Health", "Work") and the Goals page groups by it. A group exists because a goal carries its name. Your existing focus areas were converted to group labels automatically, on every Mac.
• The Plan tab is now the Goals tab, and it shows one group at a time rather than every goal at once.
• A goal is completed or it is not. The four working statuses collapsed to one honest pair. Progress lives in the goal's tasks, and the weekly AI review is where you get a read on whether it is actually moving. Feeling stuck opens a conversation instead of setting a label.
• Scheduled prompts are called routines. Same feature, a word people actually use. The Routines page is routines only and is now a sortable, filterable, searchable table. Prompts you run by hand rather than on a schedule went back to Notes as a saved prompt template.
• Goal, task and routine details are quiet property lists. Labels in a gutter, hairline rows, fields that reveal themselves as inputs when you hover. Detail pages read as one column now, with the heading at the top of it.
• The assistant finishes what it starts. Multi-step tasks now plan in a checked shape, replan around a step that failed rather than stopping, verify against the goal you actually asked for rather than against the steps it invented, and always deliver something at the end. Tasks can read attached images and files, keep their card when you navigate away, and carry the conversation they came from.
• One search in the app. Notes lost its own search box; Cmd+K is the search.
• Task lists fold completed rows behind a collapsed toggle, and tasks with no time say "anytime" instead of showing a made-up hour.
• A task's goal is a link. The goal name on a task opens it, and the × beside it is how you change it.
• Settings permission rows are one line each, and the Automations page rows are compact with their actions on hover.
• The composer mode chip is gone. There is one assistant, and it decides how to handle what you ask.
• Portfolio remembers how you sorted a table, and holdings show day-change percent.
• The home feed lost its preferences button; Unread and All is the whole control.

Fixed

• The assistant no longer says a task is running when it is not.
• Email bundles no longer count rolled-up messages as unread, so the Email card's count matches what you see when you open it.
• A goal renamed anywhere carries its review routine along, instead of leaving an orphaned routine pointing at the old name.
• Long search queries from a task step are rejected rather than failing silently, and a step that only narrates what it would do no longer counts as done.

This release is about the assistant finishing real work. It can now run things while you are away, read your actual documents instead of guessing from filenames, save a procedure that worked and replay it, and undo what it changed. Settings got simpler, moving between apps got faster, and a serious sync bug that could lose edits between Macs is fixed.

Added

• The assistant works while you are away. Ask it to do something on a schedule ("every weekday at 7, check my portfolio against the news") or when something happens (a certain email arrives, a file changes) and it becomes an automation. It runs in the background, notifies you when it starts and finishes, and if it needs your OK partway through it waits and asks when you are back instead of guessing. The new Automations page shows everything that is armed, what ran, and what is waiting on you.
• Permissions with limits. A standing permission can now carry bounds: so many uses a day, an end date, and exceptions. The assistant can hold a standing permission to send email, say, and still ask before writing to someone brand-new. When a limit runs out it asks instead of acting; nothing is ever a silent yes. Review and revoke it all under Standing permissions on the Automations page.
• It reads your documents. PDFs (including scans, read on your Mac), Excel spreadsheets, Word documents, and images: the assistant now opens the file and reads what is inside, whether it found it in your folders or on a web page. "What did this statement charge me?" is answered from the statement, not the filename.
• Recipes. When a task finishes cleanly, you can save it as a recipe. Next time, the assistant replays the steps that worked instead of figuring everything out from scratch: faster, and far more predictable. Saved recipes are listed in Settings.
• Undo for the assistant. Everything the assistant changes is recorded. "Undo this turn" under a chat answer, or "Undo changes" on a finished task, puts things back. Anything you edited yourself in the meantime is left alone and reported rather than overwritten, and actions that cannot be taken back (a sent email) say so up front.
• Your portfolio reviews itself. On weekdays, three short notes appear in your home feed, written from your real positions and prices: a pre-market review around 7:30, a midday check around 12:30 that only speaks up when something is moving, and a market-close review around 4:30 that says how the day went and whether you are still on plan. Edit the times or turn any of them off in Prompts.
• Connections you can add in one click. Settings now offers ready-made connections for the assistant: Browser, DeepWiki (ask about any public GitHub project), Context7 (current developer docs), and GitHub (your issues and pull requests, paste one token). Each one is tested the moment you add it, and problems are explained in plain words.
• Ctrl+Tab switches between recent apps, like browser tabs. Tap once to bounce back to the app you were just in; hold Ctrl and keep tapping to walk further back. A titlebar button opens the same strip with the mouse.
• An app switcher in the titlebar. The grid button beside the Anjadhe wordmark lists every app; click one to open it, or Customize to choose what shows where.
• Light, Dark, or System theme in Settings, with System following your Mac.
• Answers link to what they made. Files and folders the assistant wrote, prompts it scheduled, and records it touched appear as small buttons under the answer: click to open the record or reveal the file in Finder.
• The terminal got friendlier. Answers render as formatted text, /help, /model, /clear and /undo work as commands, and if sync is locked the terminal asks for your passphrase instead of failing.

Changed

• Settings is flatter and easier to find things in. Each setting has its own page, every page fits one consistent width, and Cmd+K now takes you straight to a specific setting by name.
• Portfolio is calmer. Your accounts sit in a left nav; click one and the whole page scopes to it. Holdings are one simple row per position. The analysis moved into conversation with the assistant, which now has the full picture.

Fixed

• Edits are no longer lost between Macs. With two Macs open, one could silently discard the other's work: portfolio transactions logged in the morning, note edits, prompt schedules. Portfolio, Notes, and prompt run-times now merge record by record, so both Macs keep what they did.
• Scheduled prompts fire once, at their scheduled time. A daily 7:00 prompt could run three times in a day, or fire at odd hours, when two Macs disagreed about its last run. The assistant can also help with this now: ask it "why did my review run at 4pm?" and it can look at the prompt and fix the schedule.
• Reading a feed post now sticks across Macs, and post bodies use the full width of the column.
• Opening a note, task, or goal from Cmd+K opens that record, not just the app it lives in.
• Plan's "Other goals" section now shows only goals not filed under a focus area, instead of every goal. The same underlying bug inflated a few counts elsewhere, including the weekly review's completed count.
• Buttons on a detail page's suggestion chip no longer get cut off.

The home page and the way you move around the app both changed in this release. The left sidebar is gone, Cmd+K is the launcher, and home now shows what your apps are holding for you instead of a wall of icons.

Added

• A home page that shows what needs you: under the message box, cards from the apps themselves — overdue tasks, what is due today, unread email insights, what is left on today's calendar, and your portfolio's value and movers. The rule is that a card with nothing to say does not appear at all, so home empties itself as you work and a calm day looks calm. Tasks can be ticked off straight from home.
• Cmd+K is the launcher: press it anywhere and you get your most-used apps first, then every other one. Keep typing and your own notes, tasks, goals, journal entries, bookmarks and focus areas join the results, ranked by how well they match. Arrows move, Enter opens, Esc closes.
• A plan for your portfolio: tell the assistant what the money is for, your horizon, the drop you could sit through, and the mix and limits you want — it interviews you one topic at a time rather than handing you a form. From then on Anjadhe measures your real holdings against the plan: which sleeves have drifted outside their bands, which guardrails are broken, what the plan does not cover. Those numbers are computed by the app, not written by the AI, and a trade that puts you off plan says so at the time.
• Clear a whole bundle of mail: drilling into a bundle now has Mark all read, Mark all unread and Archive all above the list, with a count of what they act on.
• Escalating notices become one insight: a courtesy notice, then an overdue notice, then a final notice is one thing to do, not three. The newest heads the insight and the earlier ones roll up into its timeline; nothing is lost, and "Not this" splits them back out. Overdue and return-by mail is also better caught in the first place.
• A switch for after-hours prices in Portfolio, so pre-market and post-market moves are yours to show or hide.

Changed

• The left sidebar is gone. Every page gained that width back. Home is the wordmark in the middle of the titlebar, a Home button appears on the left whenever you are somewhere else, and the app list lives behind Cmd+K.
• Home reads as a page, not a stack of dashboard cards: a greeting and the date on one line, the message box under it, then your cards, then the feed — all on one left edge.
• The feed is a digest, grouped by prompt. Each background prompt is one block showing its newest result with a line or two of what it says, with older editions folded behind "Show N earlier". An Unread / All toggle switches between what still needs you and the whole history. Email insights are no longer mixed in; they have their own card above.
• News is a page of its own, and the news pane that used to sit on home is gone — one place to read news instead of two. The Morning News digest is now built from your actual topic headlines rather than a web search.
• Profiles have been removed. The Work/Personal switcher in the titlebar and its Settings page are gone, and everything you have is simply visible in each app. Anything that was filed under a second profile now shows up alongside the rest.
• The titlebar's controls are one size with one icon style, in black and white like the rest of the app.
• Journal drops its quick-capture box; search moved into the page header.
• Tasks: the task detail page drops Duplicate.
• Portfolio: Import from file is gone — attach the file in a chat and ask the assistant to add the transactions. With several accounts, the header now has an account picker rather than a row of tabs.
• Email bundle settings show what each bundle actually catches, so the description that teaches the classifier is visible while you edit it.
• Opening a message marks it read after a moment rather than instantly, so skimming under the Unread filter no longer empties the list under your cursor.

Fixed

• Browse: turning "Protect this site" off now actually turns it off. The cookie and referrer protections ignored the switch, so sign-ins, checkouts and other cross-site flows kept failing on a page you had already unprotected.
• News: stories from Google News open the real article instead of a redirect page, and a story you revisit is no longer re-summarized from scratch.
• A reminder series no longer creates a duplicate task per reminder.
• AI-written text now renders with the same typography everywhere it appears — chat, feed posts, and news summaries.

Added

• Options in Portfolio: track long calls and puts alongside stocks. The transaction editor has a Stock/Option toggle with underlying, call/put, strike and expiration; holdings and detail views show a readable label (AAPL $250 Call 12/18/26) with an expiry chip, and the 100x contract multiplier is applied to value, cash and P&L. Long positions only for now.
• Account tabs in Portfolio: an All Accounts tab plus one tab per account under the header, so accounts switch directly instead of going back through the dashboard. Holdings also cluster by symbol, so several lots of the same stock (or several contracts on the same underlying) fold into one row.
• Email search that finds things: words match in any order across sender, subject, snippet and locally synced message bodies, with one-letter typo tolerance, accent folding, "quoted phrases", and Gmail-style narrowing (from:, to:, subject:, is:unread).
• A reading pane in Email: a message opens beside the list instead of replacing it, with a header toggle for a full-width list and a breadcrumb that names where you are. Bundles drill in rather than expanding in place, and bundle cards read icon, name and date, then who has been writing lately.
• Save a news story: the article reader has a Save button, and the topics rail has a Saved list that keeps stories after the headline drops out of the feed. Your saved list syncs between your Macs.
• Rails you can collapse: News gains a topics rail and Calendar an accounts rail, both behind a hamburger beside the title and remembered per Mac. Picking a calendar account scopes the grid to just that account.
• The assistant can reach your mail properly: it searches with the same engine the app uses (bodies included), reports how far back each account is synced instead of presenting a partial window as full history, can fetch older mail from Gmail once you confirm the timeframe, and can extract structured data across many messages in one pass ("list every subscription charge this year").
• After-hours prices in Portfolio, from a batched quote that carries the pre-market and post-market moves.

Changed

• Email is a full-height workspace: the page itself no longer scrolls, and the sidebar, list, message and insights panes each scroll on their own. Search and the Unread/All filter moved up into the header row, rows are one tight line, and unread reads as bolder ink instead of a shaded background. Starring is gone from the interface (Gmail stars still sync as data).
• The mail list stays fast into the thousands of messages: rows draw a page at a time and widen as you scroll.
• One page geometry everywhere: every page now uses the same content band, so the left rail stops sliding sideways when you enter or leave Email. Page titles all start at the same place, and the header buttons (help, settings, search, actions) sit in one order across apps.
• Settings > AI Models: calmer cards with a status dot and a Default chip, and "Add a model" opens straight into the local catalog sized to your Mac's RAM, with your own server, OpenAI and Anthropic behind "More options".
• The Labels and Accounts sidebar in Email collapses behind a hamburger.
• Portfolio's "Ask about this" buttons start a fresh conversation instead of adding to the last one.
• Anjadhe's opt-in analytics and the mobile-sync relay now run on Anjadhe Connect (api.anjadhe.com) rather than Cloudflare. Both stay keyless and carry no personal data.

Fixed

• Calendar put timed events on the wrong line: an 8:30 meeting drew at 8:00. Events now sit at their actual start minute, and titles are one clean line instead of being sliced through the middle.
• A damaged email search index could fail a mail write (marking a message unread) instead of just failing the search.
• Email: background sync no longer leaves "Load older emails" stuck on "Loading...".
• The assistant no longer mistakes a spread-out repeat for a stuck loop and ends a long research turn early.
• Web Browser: Home actually unloads the page, and audio from a page you navigated away from now stops.
• Portfolio: day change for shares bought today is measured from your purchase price instead of yesterday's close, so a morning dip you bought after is no longer shown as your loss.
• Email transaction extraction understands per-contract option pricing, and Create Transaction preselects the right portfolio account.
• The assistant's email tool returns the stored message body.
• Region-specific examples in the app's copy are now neutral or US-based.

Added

• News: a full news page for the topics you follow. An editorial front page (lead story, a Latest rail, a section per topic), a reader that fetches and summarizes articles on your Mac, a "Catch me up" digest, and upcoming dates spotted in articles offered as "Add to Schedule". The x on any story means "show fewer like this" and ranking learns from it; stories boosted toward your interests carry a visible "For you" label. Headlines come straight from news feeds, never written by the model, and your clicks and ranking stay on this Mac. News shares the Discover pane's switch and topic list (Settings > Appearance, plus web search enabled).
• Weather and local news: add an optional location in the Discover topics dialog and the pane shows today's weather (from Open-Meteo) and headlines for your area.
• The anjadhe terminal command: Settings > AI Assistant > Terminal Access installs a command that talks to the running app. Bare anjadhe opens an interactive session; anjadhe ask "question" and anjadhe task "goal" are one-shots. Same brain and same permissions, with approvals shown right in the terminal; the transcript lands in a "Terminal" conversation that stays on this Mac.
• Images in chat: attach images with the + button or drag and drop when your model can read them. Vision-capable models carry a "Reads images" badge in Settings; screenshots the assistant takes with its tools now reach vision models too, and sent thumbnails expand to a lightbox on click.
• Task mode on the composer: the mode chip now cycles agent, chatbot, and task. In task mode your message becomes a long-running job: planned into steps, approved by you, run with tools under safeguards, and reported back.
• System control for the assistant: Apple Shortcuts tools, AppleScript recipes for controlling apps, and background processes (start, poll, and stop long-running commands). Every script still asks permission first.
• One web search switch: an explicit "Enable web search" master toggle in Settings > AI Assistant > Web Search. Off means nothing in Anjadhe sends queries to the web, and every feature honors it.
• Notes tags: tag pills on a note are now clickable filters, and a "Find a tag" box in the sidebar narrows the tag list as you type.

Changed

• The assistant panel slides in over the page as a drawer instead of pushing your content aside.
• Article links from News and Discover open in the built-in browser's reader view by default.
• Anjadhe Connect now identifies your install with a random id instead of one derived from the Mac's hostname; existing installs migrate automatically, and the Connect card shows the id's anonymous hash for support.
• The in-app help guide was refreshed for all of the above and now matches the current Settings layout.

Fixed

• Security hardening: closed a shell-approval bypass via AppleScript and tightened how image data URLs are rendered.
• MCP tool servers that need npx (like the browser) now start correctly when Anjadhe is launched from Finder, and run_command finds Homebrew and nvm tools in the same situation.
• Vision model downloads use the correct file names with sha256 pins, and vision adapter files no longer appear as installable models.
• Email compose: the AI Assist panel could sit below the window edge.
• Sharing your model on the network: a watchdog now keeps the shared server alive.

Fixed

• Email insight notifications match what's in the app. The "New insight from..." notification used to fire the moment a new email was flagged for analysis, before the AI had looked at it, so emails the AI then judged not worth surfacing produced notifications with no insight behind them. The notification now fires only when an insight is actually saved, so clicking through always finds the matching card in the Insights list and home feed. It arrives a little later than before, once the analysis has really run.

Added

• Name your assistant: first-run setup now ends with an optional step to give the assistant a name, and existing installs get a small, dismissible "Your assistant can have a name" line on the home page and in the assistant view. You can also set, change, or remove the name anytime from the Name field at the top of Settings, under AI. The name replaces the "AI Assistant" label everywhere it appears — the sidebar, the chat header, and the "Ask Anjadhe anything" box — and the assistant knows its name and answers to it. It syncs to your other Macs. Skipping is remembered, so you're only asked once; leave the field blank to keep the generic label.

Added

• Wellness, a new built-in app: one place to log everything about your health. Blood pressure and pulse, resting heart rate, blood glucose, SpO2, temperature, weight, activities with heart-rate data, steps, meals, water (with a one-click quick add), sleep, mood and energy, medications, and symptoms. Blood pressure and glucose readings show how long after your last meal and activity they were taken, so the number carries its context. Trend charts cover every metric with 2-week to all-time ranges; a Units button switches lb/kg, mg/dL/mmol/L, F/C, and oz/ml. If you built the earlier blood-pressure tracker as a custom app, Wellness offers a one-time import of all its data. The assistant can log entries ("log my bp, 122 over 79"), list them, and summarize trends.
• Discover pane (preview, off by default): an optional right-hand pane on the home page with fresh links for topics you pick. Turn it on in Settings, under Appearance; it also needs web search enabled. Off unless you enable it, on new and existing installs alike.
• Wide monitors use their width: on large windows the home page arranges the rail, feed, and Discover side by side instead of one narrow column.
• Feed posts from background prompts have read/unread state, and one "Hide read items" preference now covers the whole feed.

Changed

• Email opens inline: clicking a message keeps the mailbox sidebar in place instead of taking over the whole page, and the viewer's action buttons are quieter ghost chips.
• Help no longer opens itself the first time you visit an app.

Fixed

• Background email refresh no longer resurrects the email view over whatever app you were using.

Added

• How Anjadhe works, in the app: the About view now tells the full story from the website (a canvas of apps, your data on your own SSD, the privacy promise, and only then the AI), so you can read it without leaving the app.

Changed

• Email insight cards read like inbox rows: the sender leads the card, the subject is a real title instead of an uppercase chip, the AI summary sits quietly underneath, and each action item is its own row with a due-date chip that turns red once the date has passed.
• Feed cards show their source at a glance: a subtle accent stripe on the left edge marks email insights (teal) and prompt results (purple); posts from Anjadhe stay neutral.
• Dates read like a person wrote them: insight titles, summaries, and action items show "Jul 31" instead of "2026-07-31", everywhere insights appear.
• The setup checklist got out of the way: the "Set up more of Anjadhe" banner above the chat box is now a quiet pill in the home page's bottom corner. Click it to continue, or dismiss it with the ×.
• "Push to today" moved next to the Overdue heading on the Actions and Tasks pages, instead of sitting in the far corner.
• Slimmer home feed header: prompt creation lives on the Prompts page (its + New Prompt button); the feed header keeps preferences, Clear all, and Manage prompts.
• Notes: + New Note is a plain button again: every new note starts blank, and the editor's Template control switches it in place. The dropdown caret is gone.
• Notes: Show-on-Home removed: the home page no longer renders pinned notes, so the house-icon toggle and the On Home filter did nothing and are gone.

Fixed

• Email insights appear on home from first launch: previously the feed stayed empty until you visited the Email app once. Mail sync and analysis now start at launch, and newly analyzed insights appear on the home feed live, without navigating or refreshing.
• Background AI waits for a configured model: email analysis and bundle classification used to fire doomed calls when no model was set, littering AI Activity with errors. They now wait quietly and catch up the moment you pick a model.

Added

• Starter prompts out of the box — every install now comes with four scheduled background prompts: Daily Briefing (your tasks and calendar each morning), News Digest, Daily Motivation, and Weekly Reflection. They're ordinary prompts — edit, reschedule, or delete them on the Prompts page.
• Email insights on the home feed — unread insights (renewals, bills, appointments) appear as feed cards alongside prompt results. The × marks one read; clicking opens its new detail page.
• Insight detail page — clicking an insight (in Email → AI Insights or on the home feed) opens a full page: summary, action items linked to their tasks, feedback buttons, and the underlying email. Opening marks it read.
• Feed preferences — the sliders button in the feed header chooses what the feed shows: prompt results, email insights, and whether read insights stay hidden.
• Per-page help — every feature page has a small ? in its header that opens that page's guide in place; it also shows itself once on your first visit. New Help articles for Browse and AI Activity.

Changed

• Bookmarks open in the built-in browser — clicking a bookmark opens the page directly (with a "Back to Bookmarks" strip); the i button on a card opens its details.
• Notes inherit the selected tag — a note created inside a tag-filtered list picks up that tag (visible as a removable pill), so it no longer vanishes from the list it was created in.
• "Run now" queues instead of bouncing — a prompt run clicked while another run is in progress waits its turn and runs next, in click order.
• AI Activity tells the truth — background prompt runs that use your personal context now show as "Background prompt · automatic" (they used to masquerade as "Assistant chat" at interactive priority), and the status line distinguishes running from waiting ("1 AI task running, 2 waiting").
• Visible self-repair — when the assistant retries a defective answer (empty response, promised-but-unrun tool), the chat shows a "↻ automatic retry" step chip, and the internal retry nudges identify themselves so the model's thinking no longer narrates a message you never sent.
• Friendlier failure cards — a failed background-prompt run posts a quiet status note ("The model came up empty this time — it will try again…") instead of a red engine error; empty responses are retried once first.

Fixed

• Server/cloud API key reached only chat — email insights, bundling, compose, goal updates, memory, task mode, Maker, and background prompts now carry the default model entry's API key; they used to fail with "Invalid api key" on a server or cloud brain while chat worked.
• Double app initialization — navigating to an app re-ran its init a second time via the app's own URL-hash write; self-caused hash changes are now ignored (browser back/forward still works).
• Scheduled prompts with no model — the scheduler now skips quietly until a model is selected instead of posting an error card per prompt.
• Help/settings header buttons sized to match their sibling buttons.

Fixed

• Built-in web search recovers on its own if the service stops recognizing this Mac's access key (it now discards the stale key, gets a fresh one, and retries — previously every search failed with "Unknown API key" until the key was refreshed from Settings). Monthly usage is unaffected by the refresh.

Added

• Web search is built in. The assistant can search the web out of the box — no account, no API key, nothing to configure. Searches route through Anjadhe Connect, a small relay that forwards the query — and only the query — to a search provider: it keeps a count of searches, never what you searched, your Mac identifies itself with a random token rather than an account, and the provider sees queries arriving from the relay, not from you. 300 searches a month included; the relay's source code is public so all of this can be checked. Your own Tavily or Brave key still works exactly as before (Settings › AI Assistant › Web Search) and keeps Anjadhe out of the loop entirely.
• Prompts get their own page, sectioned by scheduled vs. on demand. Prompt notes no longer clutter Notes, and the Notes sidebar drops the Prompt Feed filter.
• Email: per-account scoping. The sidebar's Accounts switcher narrows the Email view to one account when you have several connected.
• Actions: the group labels in the left nav now act as filters too.

Changed

• The assistant handles multi-step research more reliably: big fan-out questions run in task mode, real results carry between steps, and it no longer answers with a promise to do the work later.
• Setup and Settings copy updated for built-in search (the web-search help article was rewritten around it).

Added

• Actions filters now combine: pick a date window (Today, Tomorrow, This Week, This Month, Later — or the new All time) and a focus area at the same time, and the list narrows to both. Pills above the list show the active filters; the × on a pill clears just that one. A new Dates header labels the date filters like Focus Areas.
• A No focus area view in the Actions nav shows every task filed under nothing — the tasks that were invisible in every focus-area view.
• Completed tasks show under every filter in a collapsed "N done" section — Tomorrow, This Week, This Month, Later, All time, focus areas, and No focus area (Today already had it).
• Task rows have a ⋯ menu (also on right-click): Open, Set date, Rename, Delete — quick edits without leaving the list. Same menu the Plan task lists already had.
• Quick-add takes its date from the page you're on: adding a task on the Tomorrow page schedules it for tomorrow, on Later it lands in the undated backlog. A date typed in the text still wins.
• Portfolio: a Holding transaction type to enter existing positions directly, and a more compact transaction editor.

Changed

• Task edits save automatically — the Save button is gone from existing tasks; only creating a new task asks you to save. Leaving the task page (Back to list, Esc, or any nav click) never discards typed changes.
• The Actions and Plan pages drop their breadcrumbs and big headings — the selection pills and the left nav say what you're looking at, once. The task detail gets a "Back to list" link and closes on Esc.
• The old standalone Tasks and Goals pages are retired; every path into them (calendar events, email insights, goal links in Notes/Bookmarks, Pomodoro) now lands on the Actions hub or Plan's goal detail.
• "Open in Plan" is now "Edit Plan"; the focus-area view lost its redundant All/Today/This Week chips (the nav's date filters do that now, and they combine).
• Tighter Actions lists (more tasks per screen), wider Tasks | Plan tabs, compact task-detail buttons, and the "Restart & update" flow now asks before interrupting AI work that's still running.

Fixed

• The assistant now finds schedule tasks reliably: telling it you did something searches your schedule (any date) instead of trusting a truncated list — long schedules no longer hide tasks from it — and when a task covers several things and you finished only one, it asks about the rest instead of marking the whole task complete.
• Clicking the already-active date filter while a task was open silently cleared the filter and dumped you on the wrong list — it now just takes you back.
• Background AI classifiers no longer spin in retry loops (model thinking disabled, decode capped).

Added

• The AI engine's heartbeat lives in the titlebar now: a pulse icon that animates with an amber dot whenever AI work is running — chat, email insights, background prompts, or a model loading. Click it any time to open the AI Activity page and see exactly what's happening.

Changed

• The browser's reader view strips much more page clutter — menus, social share bars, "related stories" blocks, newsletter prompts, and link-heavy navigation that carries no class name at all — while keeping the article and its inline links intact. Body text now reads like Notes and Journal: system font at reading size, with the serif reserved for the title and headings.
• The AI Activity status card says where your model actually runs: with a server or API model as your brain, it now reads "Your model runs on your server" instead of implying the current model is whatever happens to sit in local memory — a leftover local model is called out as exactly that.

Fixed

• The assistant's status dot no longer flips to "Busy" right after "Preparing…" during startup — the whole warm-up now reads "Warming up…", and "Busy" is reserved for real work like email insights or a chat running in another conversation.

Added

• AI Activity (System group in the sidebar): a live view of what the AI engine is doing on this Mac — current and recent work in plain words, not logs. Click any row for the details: what ran, who started it, which model, how big the request was, how long it took, with a jump into the full LLM logs. If the GPU spins up while you're not chatting, this page shows why.
• The sidebar can now be minimized: hover it and click the chevron to collapse to an icon strip on every page. Hovering the strip still flies it open; the choice is remembered.

Changed

• llama.cpp is now the one local AI engine — Ollama support was removed. Model entries carry over automatically.
• 16 GB (and smaller) Macs run local AI strictly one request at a time: parallel generations were halving each other's speed. Your chat takes priority, and background work waits for a pause in the conversation instead of squeezing between your messages.
• Email insights are leaner: responses are JSON-constrained and capped (a verbose model could previously decode for many minutes per email), and every judged email is remembered — including "nothing here" — so the same mail is never analyzed twice.
• The assistant's memory-notes pass runs far less often and stays out of your way; the conversation-goal pass is skipped entirely on 16 GB Macs.
• Recurring prompts are now called background prompts everywhere.
• First-run setup now recommends Qwen3.5 9B on 16 GB Macs; Gemma 4 12B and E4B remain one click away.
• Notes: automatic tag suggestions were removed (manual tags unchanged).
• Quieter chrome: the titlebar quick-add buttons are gone, and detail titles in Actions and chat headings sit a calmer size.

Fixed

• A context-size mismatch made llama-server reload the whole model on nearly every call on 17–32 GB Macs — AI felt minutes-slow until it was found. Both "auto" tables now agree for every RAM size.
• Switching local models no longer throws an uncaught error.
• The assistant's status dot is honest now: it shows Busy when the engine is mid-way through background work, and a message sent then says what it's waiting for instead of showing anonymous dots.
• Memory wiki pages: the editor expands to fit long pages instead of clipping them at a fixed height.

Added

• Notes is now a three-pane notebook: filters, list, and the note itself — always open, always editable, autosaved. Quick capture at the top of the list, full-text search as you type, and [[wiki links]] between notes with backlinks and tag-based related notes. With a local model configured, quiet tag suggestions appear under the note.
• The assistant's memory grew from fixed sections into a wiki of topic pages: it starts pages for people, projects, and themes on its own, and reads the right page back when a conversation needs it. Browse and edit everything via memory in the chat header.
• Actions: search across tasks, goals, and focus areas from the header — Cmd/Ctrl+F focuses it, arrow keys walk the results.
• Ask the assistant for a recurring prompt in plain chat ("every morning, show me…") and it sets up the scheduled prompt for you — you approve it before it's created.

Fixed

• Dark mode: invisible modal text and a handful of undefined colors.
• Chat no longer looks idle while the assistant works silently through tools mid-turn — it names what it's doing.
• Actions: a pending filing suggestion clears once you file the task yourself.
• Smoother 12B-model behavior on 16 GB Macs: background jobs queue instead of piling on, and the model unloads under memory pressure.

Changed

• The project moved to its own GitHub organization: github.com/Anjadhe/Anjadhe. Existing installs keep updating seamlessly.
• Scheduled prompt runs know no one is watching — they finish the job instead of asking clarifying questions into the void.

Added

• Links open in Anjadhe's own browser, iOS-style: click a link in the feed, a chat, a note, a journal entry, or a bookmark and it opens in an in-app browser tab with a slim "‹ Back" strip that returns you exactly where you were — the post you were reading, the chat, the note.
• A feed post and its prompt now link to each other: the post's meta line opens the prompt note, and a prompt note shows its recent runs as compact feed cards.
• Web search: a per-provider "Parallel searches" setting.

Fixed

• Links in feed posts did nothing when clicked — they now open reliably (in the in-app browser).
• With a server as your AI brain, background work (email insights, goal updates, feed prompts) no longer probes local Ollama first — it goes straight to your server, and the console stops filling with connection-refused errors.
• Task lists under goals and focus areas show abandoned tasks as resolved instead of pending.
• "Suggest tasks" no longer lands every suggestion on the same date.

Changed

• The About page says the same things in half the words.

Added

• When an update has finished downloading, the home page now shows an "Update ready" card with a Restart & update button — a more visible companion to the titlebar pill. Dismiss it and it stays gone for that version; either way, the update still installs automatically the next time you quit.

Added

• The About page links to Anjadhe's public source code on GitHub — you don't have to take the privacy claims on faith.

Changed

• Abandoned tasks now count as done everywhere. Marking a task abandoned (deliberately not doing it) resolves it across the whole app, exactly like completing: it checks off in every task list, disappears from the calendar, stops its reminders, and the assistant treats it as finished instead of pending or overdue. The assistant can also record an abandonment when you tell it you're not doing something.
• A lighter setup checklist. Connecting Google is one step — turning your inbox into tasks happens automatically after you connect, so it's no longer a separate chore. The web-search step is gone (add a search key any time in Settings › AI), and "Add a frontier model" now appears only on Macs under 32 GB, where the local model can't carry the assistant on its own.

Fixed

• The home-page setup nudge could claim "4 of 5 done" while the Settings checklist showed everything complete; the two now always agree.

Fixed

• Auto-update actually works now: releases ship the ZIP the macOS updater installs from. Installs of alpha.4 and alpha.5 detected new versions but silently failed to download them — this release is the first they can update to; from here on, updates apply on quit.

A small follow-up to alpha.4.

Added

• A lighter local model option in the catalog for Macs with less memory.

Changed

• The feed and the post reading page now share one, slightly smaller type scale.
• Copy touch-ups across the app.

Fixed

• Releases publish immediately instead of landing as invisible drafts, so the in-app updater sees them right away.

This alpha rolls up several months of work and is the first release built and published through the new pipeline. The theme: one assistant that lives with your data — private by default, honest about where everything runs.

Added

• A real assistant. One chat that reads your tasks, goals, notes, journal, email insights, and calendar — and acts: creates tasks and notes, files action items, searches the web (bring a free Tavily key), reads pages, and takes on long jobs as planned, approvable tasks. It keeps durable memory you can read and edit, shows a Sources row built from what it actually did, and a per-chat chip switches to plain chatbot mode.
• Artifacts. Ask for a document, a slide deck, or a small interactive page — the assistant builds it and it exports to PDF. Maker does the same from a single prompt with its own web research.
• Bring your own frontier model. OpenAI and Anthropic can be added as model entries with your own API key — an explicit per-model choice, never a fallback. Local stays the default; a server you own remains supported.
• Chat attachments. Drop text files or PDFs into the composer; parsing happens on your Mac.
• Build your own apps. App Studio builds simple apps from a description, and any coding agent pointed at ~/Anjadhe/apps/ can build richer ones — new apps get storage, encrypted backup, sync, and the assistant's attention automatically.
• Sync passphrase. Optionally wrap the sync encryption key in a passphrase (Settings → Storage & Backup) so your iCloud alone can't read your data.

Changed

• First run, redesigned. The welcome draws the privacy boundary — your data inside a frame labeled "Your Mac" — then downloads one good local model (Gemma 4 12B QAT) through the app's built-in llama.cpp engine; no separate installs. You land in the assistant, with a waking-up page while the model loads the first time.
• Actions is now two tabs. Tasks opens to Today — one finishable list, with Tomorrow, This Week, This Month, Later, and each focus area a click away — and Plan holds focus areas with their goals. The assistant files loose items, suggests goal breakdowns, and runs a five-minute weekly review.
• Email insights. Bundles for the noise, summaries for what matters, and action items that land in Actions on the right day — analysis on your Mac.
• Setup checklist became a discovery panel. A quiet "Learn what else you can do" pill replaces the auto-opening checklist; it now also covers the web-search key and bring-your-own-model options.
• Copy is honest everywhere. Present-tense privacy claims only, and the same plain language in the app and on anjadhe.com.

Fixed

• Fresh installs now show the downloaded model in the composer's model menu and the Settings model library (the entry list no longer initializes empty during setup).
• The packaged app includes the llama.cpp manager (previous packaged builds would have failed at launch) and the universal build no longer breaks on a stray native module.
• Dozens of smaller fixes: notes title focus ring, streamed reasoning contrast, breadcrumb spacing, OAuth result pages styled to match.

Third alpha. The theme is honesty about what runs where. The model picker now shows accurate sizes, sets RAM requirements that reflect what actually feels good (not just what technically loads), and — for Macs that can't run local AI well — offers Claude as a one-click cloud fallback with automatic PII stripping.

What's new since alpha.2

• Recalibrated local model catalog. Every entry verified against its ollama.com page. The big miss: gemma4:latest was previously shown as 5.5 GB (it's actually 9.6 GB). Most others were off by a few hundred MB. More importantly, the minimum-RAM floor is now "runs well" — not "barely loads" — so a 16 GB Mac is no longer offered Gemma 4 as a default that feels sluggish. Model tags are pinned (no :latest aliases) so the catalog stops drifting silently when Ollama re-points them. Added the full Gemma 4 variant lineup (e2b, e4b, 26b, 31b) and Gemma 3 27B for larger Macs.

• Cloud fallback for low-RAM Macs. If your Mac has less unified memory than any local model needs to run well, the first-run wizard now offers to set up Claude (Anthropic's cloud) instead, with automatic PII stripping before your messages leave your device. Or you can skip AI entirely and configure later in Settings. Previously this path showed a confusing "Could not load the model catalog" message and left you stuck.

• Fixed first-run on fresh installs. Two files that should have been bundled into the app (the model catalog remote-config.json and the app icon used by the About view) weren't in the asar, so clean installs of alpha.2 couldn't populate the model picker or render the About hero. Both are bundled now.

• Remote (Claude) path now actually works end-to-end. Previous builds could silently return empty replies on API errors (bad key, wrong model) and the assistant would make up plausible- sounding explanations for the failure. HTTP errors are now surfaced properly. The PII anonymizer also got a major upgrade: it uses reversible per-chat tokens instead of a blunt [email] placeholder, so tools that need a real value — like sending an email to a real address — work against the remote model. The privacy guarantee is unchanged; PII still never leaves your Mac.

• Assistant understands your portfolio. The agent now has tools for current price, value, gain/loss, day change, allocation, and single-ticker drill-down. Ask "what's my net worth" or "how's my NVDA doing" and you'll get grounded answers with live prices and a refresh option when the cache is stale.

• Web search (optional, bring-your-own-key). Add a Tavily key in Settings and the assistant can look up facts it doesn't know — current events, news, product specs. Search queries go to Tavily; your local data does not. The free tier is plenty for personal use.

• Accurate overdue-task count. The assistant was previously inflating "N overdue tasks" by counting one-time tasks you'd already completed on prior days. The count now matches what the Schedule app shows.

• Chat bubbles render links. Markdown [text](url) and bare URLs are clickable and open in your default browser. Numbered lists no longer render as a sequence of "1." items.

Install

Download Anjadhe.dmg below, open the DMG, drag Anjadhe to your Applications folder, launch from Applications. Universal binary — runs natively on both Apple Silicon and Intel.

Upgrading from alpha.2? You'll see the update nudge in the titlebar shortly after launch; click Install to restart on alpha.3. Or replace the app in Applications manually. Your data, conversations, and settings are preserved either way.

Upgrading from alpha.1? Alpha.1 shipped before the auto-updater existed, so you'll need to install alpha.3 manually this one time (download the DMG and replace the app in Applications). From alpha.3 onward, future releases will come through auto-update.

Known limitations

• macOS only. Tested on macOS Tahoe; earlier versions are untested.
• Local AI requires Ollama with a supported open-weight model. Response speed scales with your hardware — more unified memory and a beefier chip means faster turnaround. Macs with less than 16 GB of unified memory are offered the Claude cloud fallback instead of a local model.
• This is still an early alpha. Please file issues in the issues tab of this repo.

Privacy

Your data stays on your device. No telemetry, no tracking, no cloud servers. Personal data sent to the optional Claude integration is automatically anonymized before leaving your machine. See anjadhe.com for details.

A note about the auto-attached "Source code" archives

GitHub automatically attaches "Source code (zip)" and "Source code (tar.gz)" links to every release, and there's no way to hide them. On this release those archives contain only this repo's .gitignore — the actual Anjadhe application source lives in a separate private repository. See the repository README for more.

Second alpha. Two headline changes: when you open a chat with the AI assistant, it already knows your focus areas, active goals, today's schedule, what's overdue, the unread action items sitting in your inbox, and your latest journal entry — no more "let me check…" before the first response. And from this release forward, new versions will auto-update.

What's new since alpha.1

• Daily briefing for the AI assistant. Every conversation starts with a snapshot of your focus areas, in-progress goals, today's tasks, real overdue items, unread email action items, and your most recent journal entry. The assistant uses this as starting context, so asking "what should I focus on today?" gets an immediately specific answer instead of a generic tool hunt.

• Human-friendly date formatting. Overdue tasks, journal entries, and similar references appear as "yesterday", "3 days ago", "April 2" — never raw ISO strings.

• Auto-updates. New versions download in the background and nudge you to install with a one-click restart. You can also trigger a manual check from Anjadhe → Check for Updates… in the menu bar.

  ⚠️ One-time manual upgrade for alpha.1 users. If you installed alpha.1, you'll need to install alpha.2 manually this one time by downloading it from below and replacing the app in your Applications folder. From alpha.2 onward, future releases will come through auto-update automatically. This is a fundamental limitation of bolting auto-update onto a build that didn't have it — there's no way around it for the first hop.

• Refreshed About view. The in-app About screen now uses a hero-style layout matching the website, with the new positioning: a privacy-first productivity app with a built-in local AI assistant that lives with your data.

• Small fixes. Tasks with empty titles in the overdue list no longer confuse the assistant. Budget app hidden from marketing surfaces until it's wired into the AI assistant.

Install

Download Anjadhe.dmg below, open the DMG, drag Anjadhe to your Applications folder, launch from Applications. Universal binary — runs natively on both Apple Silicon and Intel.

Upgrading from alpha.1? Replace the app in Applications. Your data, conversations, and settings are preserved.

Known limitations

• macOS only. Tested on macOS Tahoe; earlier versions are untested.
• Local AI requires Ollama with a Gemma 4 model (or another open-weight model of similar capability). The app prompts you on first run if Ollama isn't installed. You'll want a Mac with enough unified memory to run it comfortably — response speed scales with your hardware.
• This is still an early alpha. Please file issues in the issues tab of this repo.

Privacy

Your data stays on your device. No telemetry, no tracking, no cloud servers. Personal data sent to the optional Claude integration is automatically anonymized before leaving your machine. See anjadhe.com for details.

A note about the auto-attached "Source code" archives

GitHub automatically attaches "Source code (zip)" and "Source code (tar.gz)" links to every release, and there's no way to hide them. On this release those archives contain only this repo's .gitignore — the actual Anjadhe application source lives in a separate private repository. See the repository README for more.