MicJotAbout the beta

AI writing

AI writing is optional. Without it, rule-based formatting runs on your Mac. With it, MicJot also removes false starts, applies corrections, matches your tone, and powers Command mode and Polish.

What AI writing adds

  • Cleanup: “um”s, false starts and self-corrections (“at two, actually three” → “at three”) go; lists are formatted; the text continues what is already in the field; the tone matches the style you chose for the app.
  • Command mode: hold Control with the dictation key and say an instruction. With text selected, it rewrites the selection (“make this friendlier”, “translate to Spanish”, “turn into bullets”); without a selection, it writes new text or answers.
  • Polish: select text and press ⌃⌘P to rewrite it clearly. Leading and trailing whitespace of the selection is kept.

The switches live in Settings → Writing: Smart formatting (the rule-based path, on every dictation, with or without AI), AI cleanup and Use text near the cursor; when AI cleanup cannot run as described, one line under them says why and what changes it. The style for each kind of app is chosen on the main window's Style page.

In the private beta: your own Anthropic key, in QA builds

The builds testers receive are QA builds (MICJOT_DEV_TOOLS), and only those have a key field: under Settings → Writing → QA build only sit Anthropic API key, Claude model (defaults to claude-haiku-5-5; the field hints when an id looks wrong) and Prefer my own Anthropic key over MicJot AI. A customer build has no own-key UI anywhere, so this route exists for the beta and not for the product.

The key is stored in secrets.json in MicJot's support folder (readable only by you) and sent only to Anthropic.

What is sent, only when a key is set and an AI feature runs: the recognised transcript (or, for Command mode and Polish, your instruction and the selected text), the frontmost app's name and category, your tone and language settings, up to 200 words from your dictionary, and, when “Use text near the cursor” is on, up to 600 characters before the cursor. Over HTTPS, with your key; Anthropic's API data policy applies. Nothing is sent for excluded apps, password fields, or while Secure Event Input is on.

What happens when the AI fails. The dictation never waits on it: cleanup has a 10-second budget, and on a timeout, a network error or a rate limit the rule-based text is inserted instead, with one notice. A key error pauses AI until the key changes; three network failures or timeouts in a row pause it for five minutes. A reply that is not a cleaned transcript (an answer to the text, a preamble, a quoted or fenced copy, a different language, a drastic length change) is rejected and the rule-based text is used. A failed Command keeps your instruction in History (or on the clipboard when History is off).

Without a key, Command mode inserts your words as a dictation and says so, and Polish does nothing and says so; the Writing tab's line reads “This build has no account service, so dictations use the offline cleanup and Command mode and Polish are off.”

Cost. Anthropic bills your key. At list prices a short dictation on Claude Haiku 5.5 costs roughly $0.00014 uncached, less with the system prompt cached; this is the repository's estimate, not a measurement, and the app logs the real token counts per call.

MicJot AI, with the public release

MicJot AI runs the same features through MicJot's proxy with MicJot's key, so you need no Anthropic account and enter no key. It is included on every plan; see pricing. It does not exist in 0.3.1; this is the design:

  • You sign in once, on the first setup screen (email code, Google or Apple), and approve this Mac. The Free plan covers your first 3,000 words with every feature, and when they are used up a 7-day Pro trial is offered once, no card; Pro removes the limit.
  • The proxy sends the same data as the own-key route to Anthropic and records token counts only: never text. Anthropic deletes API data within 30 days by default, except content its safety systems flag under its Usage Policy (kept up to 2 years, classifier scores 7 years) and what the law requires; it does not train on API data by default, except feedback a user explicitly submits, which the proxy never does.
  • Each plan has a daily budget of requests and tokens, printed on the pricing page. Reaching it never stops dictation: the rule-based text is inserted with one notice until midnight UTC.
  • If the proxy is unreachable, AI is simply unavailable for that dictation; you are never signed out by an outage.

Turning it off

Switch off AI cleanup in Settings → Writing and every dictation keeps the offline rule-based cleanup. In a QA build, remove the key under Settings → Writing → QA build only, or leave it and turn off “Use text near the cursor” to send only the transcript. Private dictation, excluded apps and password fields send nothing regardless.