AnjadheWhen the model runs off this Mac
With a local model nothing you do leaves the machine. When your default model is Anjadhe Cloud, your own OpenAI/Anthropic key or a server you run, every prompt goes there: the model has to read the text to answer. Anjadhe keeps that honest in four ways.
1. Chat sends what you typed; background work sends only what you allow
What you ask in chat goes to the model exactly as you typed it, with the context you attached. Anjadhe also does work without being asked: email insights, thread status, routines, reviews. Settings › AI Assistant › Cloud Privacy has one switch per kind of data that background work may send: Email, Notes, Journal, Wellness, Portfolio, Web pages, Files. Journal and Wellness are off by default. A switched-off kind stays on this Mac: the insight sweep pauses, a routine that tries to read it is told why, and AI Activity shows a "kept on this Mac" row so nothing is skipped silently. Asking about that data in chat still works. The switches sync between your Macs; they take effect the moment you pick a model that runs elsewhere.
2. Every request that left is listed
Settings › AI Assistant › LLM Logs, filter Left this Mac: each request with its destination, the exact messages it carried, and the reply. The list is built from the same object the app sent, so it cannot differ from what went out. The model chip in the composer carries a small arrow whenever the model runs off this Mac.
3. Less goes than you might think
Before an email body leaves for background analysis, quoted history, signatures and tracking links are stripped, and background prompts are capped in size. Anjadhe Cloud requests carry no account, name or email address, only a per-install key.
4. The server is checkable, not just described
Anjadhe Cloud runs on Anjadhe Connect (api.anjadhe.com), whose code is public. Prompts and answers are never logged or stored; a test in that code sends marked text through every path and fails if it shows up in a log line, an error body or the database. What is stored: request and token counts per install per month, keyed by a hash, deleted after 400 days. The Anjadhe Cloud card in Settings › AI Assistant › Models shows the commit the server is running so you can compare it with the public repository. Requests go on to DeepInfra (deepinfra.com), the inference provider, under a no-retention agreement; DeepInfra sees Anjadhe's server, never you. Any change to what is stored is announced at least 30 days ahead.
Changelog of what the service stores
- 2026-08-25: usage counters (search and AI requests per install per month) now age out after 400 days. A canary test pins "never logged" in CI, and
/v1/versionreports the running commit.