Agent-readable docs index: /llms.txt. Full docs in one file: /llms-full.txt. Download /docs.zip to grep all markdown files locally.

Voice Calls

Join the Kimaki voice channel and talk. GPT-Live answers: it listens while it speaks, so you can interrupt it any time. For real work it hands your request to a fast backend model (gpt-6-luna), which controls your coding agents with the kimaki CLI: it lists threads, reads them, and starts new ones.
you ──voice──▶ Kimaki voice ──WebRTC──▶ GPT-Live ──request──▶ backend ──shell──▶ kimaki send ▲ │ ◀──result──────┘ session list / read └─────── speech ───────┘ text chat: transcripts, tool lines, links
A thread and a session are the same thing: each Discord thread has one OpenCode session. Say either word in a call. The CLI commands say session and take thread IDs too.

How it works

  1. Kimaki creates one Kimaki voice <machine> channel per computer, outside any category. Drag it where you want.
  2. You join it. Kimaki joins too, greets you, and listens.
  3. You ask for something. GPT-Live passes it to the backend model, which runs kimaki commands in the Kimaki data folder (~/.kimaki) and searches the web. GPT-Live says the result in its own words.
  4. For coding work it starts a thread with kimaki send --wait in the background, in the right project channel, and adds everyone in the call to it.
  5. When the thread finishes, the model gets its transcript and tells you what it did.
  6. The call ends when the last user leaves, or when you ask the model to hang up.

The text chat

Every Discord voice channel has a text chat. Kimaki uses it to show the call:
  • your words as » tommy: ... and the model's answers
  • each command the model runs, like ┣ shell kimaki session list
  • links and IDs the model posts, so it never reads them aloud
  • files from kimaki upload-to-discord, which posts here when called during a call
Type in the chat to talk to the model without speaking.

What to ask

Which threads are running now? Start a thread in the website project to fix the broken signup link. What was the last thread about the parser?

Threads report back

The model starts every thread with --wait as a background command:
kimaki send --channel <project channel> --prompt 'Fix the broken signup link' --user <you> --wait
The thread shows in the text chat at once, and you can keep talking. When the thread finishes, the same command returns its transcript. So the model gets the notification and the result together, and tells you what the thread you just started did, with no extra kimaki session read.
Other slow commands like kimaki session wait also run in the background.

Finished thread notifications

When a thread of any project finishes or fails during a call, Kimaki posts a line in the text chat with the title, the thread link and the session ID:
thread finished: Fix signup link ⋅ https://discord.com/channels/<guild>/<thread> ⋅ ses_abc123
The model gets the same notice, so it never needs to poll kimaki session list:
ThreadWhat the model does
started in the calltells you it finished
started in the call with --waitwaits for the transcript, then tells you the result once
started elsewhere (Discord, another agent)keeps it as context, mentions it when you ask

Setup: a ChatGPT subscription

Voice calls need Sign in with ChatGPT in OpenCode. Both models run on your ChatGPT subscription, with no API credits:
PartModelDoes
voicegpt-live-1-codex (GPT-Live)listens and talks, over WebRTC
backendgpt-6-luna, low reasoningruns kimaki commands, searches the web, posts links
kimaki login openai
Kimaki reads the token OpenCode stores and never refreshes it itself. If you are not signed in, or the token expired, the text chat says so and the bot does not join. API keys are not used for voice calls.
Everyone in the call is heard at once: GPT-Live gets one mix of all speakers. Only users who can use Kimaki (server owner, administrators, Manage Server, or the "Kimaki" role) are heard.

Choose the voice

The default voice is cove. The first line of a call shows it: voice call started ⋅ gpt-live ⋅ cove. The other voices are arbor, breeze, ember, juniper, maple, sol, spruce and vale.
Pass --voice when you start Kimaki:
kimaki --voice juniper # names are case-insensitive kimaki --voice default # back to cove
An unknown name stops the start with the full list of voices. The running bot keeps running. A saved voice GPT-Live does not know (for example marin from an older Kimaki) falls back to cove, and the text chat says why.

The voice is saved

Kimaki saves the voice in its SQLite database (~/.kimaki/kimaki.db). Pass the flag once; you do not need it again:
EventVoice
kimaki restartkept
bot crash, computer rebootkept
autostart when OpenCode starts (~/.kimaki/autostart.command)kept
kimaki with no --voicekept
kimaki --voice <other>replaced
The precedence is: --voice flag, then the saved voice, then cove. The autostart script does not need to change: it reads the saved voice like any other start. Each data directory (--data-dir) has its own voice.