Write-only: a key can be set or cleared, never read back. A new key is verified before it is saved, and takes effect immediately.

Mistral has no default voice, so one must always be set. Clone your own from a recording below and it becomes active immediately.

Only upload a voice you have the right to clone. Mistral stores the clone on your account; after this only the voice id is sent with each request.

There is no word budget any more — it was clipping complete sentences and announcing the cut. Brevity is the persona's job now: at most three sentences, each under fifteen words. To change it, edit the persona text.

Leave this at 0. A cap of 1500 used to be hardcoded here, and it was destroying replies: this model spends tokens on internal reasoning before it writes anything, and reasoning is billed as output — so a hard question could use the entire budget thinking and return nothing, with a red "output limit reached" warning. At 0 the parameter is not sent at all and the model stops when it is finished. The API's own ceiling is 393,216; a larger number is clamped rather than rejected. Set a value only if you want to bound cost.

A persona replaces the whole system prompt, but every one of them keeps the spoken-style rules — short sentences, no markdown, numbers written as spoken.

Neither provider gives us search for free: DeepSeek accepts a web_search tool but never executes it, and Mistral's real connector is disabled on this account. So search is our own, and when this is on the model may look something up before answering.

perplexity is the one to use: ranked results with real dates and extracted content, and it does not block automation. news is a keyless news feed that works with no setup, and remains the automatic fallback if the chosen provider fails. duckduckgo needs no key but throttles hard and is not usable.

Both token fields must be set together. Measured against this account: Tokens / page on its own does nothing at all — 2000, 8000 and 50000 all returned identical content, silently. It only takes effect once a total budget is also set, and then a single page returned 96,983 characters. Setting Tokens / page alone fills in a total for you rather than doing nothing.
Rough guide per 5-result search: 0/0 ≈ 19,000 characters, 60000/12000 more than that, both large ≈ 120,000+. Deep research runs many searches, so this is the setting that decides what it costs.

Estimates, from what has been logged. The food day is the Pacific date, because your day is: awake about 19:15 and asleep about 10:00 Tbilisi, which is 07:15–22:00 in Portland. A meal after midnight Tbilisi still belongs to the day you were awake for.

Terms the recogniser gets wrong — names, places, jargon. They bias transcription only; nothing here reaches the model.

Detection runs on the server (Silero VAD over the PCM this page streams up), so there are no thresholds here. Send now hands over immediately; Interrupt or Esc always stops it.

Every utterance is compared against my_voice.wav before it reaches the transcriber, using a speaker-embedding model on the server. Voice that is not yours — the assistant's own voice coming back through the microphone, a television, somebody else in the room — is discarded here, so it is never transcribed and never answered. The first three seconds of each utterance are used, and anything too short to judge is let through rather than thrown away. Lower the number if your own voice is ever refused; raise it to be stricter. Changing either needs a server restart.

Plays in this page when the alarm goes — the 7:35 pm wake alarm, and any one-off alarm you ask the droid for. It loops until you say you are awake (or until the droid stops it), so upload something you will actually get up to rather than something you will sleep through. An MP3, M4A, OGG or WAV; up to 32 MB.