Kanutus MCP

Model Context Protocol server for Kanutus: meetings and calls with live voice translation. Lets Claude, ChatGPT and your own agents create rooms, schedule meetings, book times on booking links, list voices, read usage and transcripts, and take part in a room as an AI agent (text in, translated voice out).

This server only talks to the public Kanutus API (https://api.kanutus.com/v1). It has no privileged access: it can do exactly what the permissions you grant allow. MIT licensed.

Use it

Claude and ChatGPT (remote connector, OAuth)

Connector URL: https://mcp.kanutus.com/mcp

• Claude: Settings → Connectors → Add custom connector → name Kanutus, paste the URL → Connect.
• ChatGPT: Settings → Apps & Connectors → Advanced → enable developer mode; then Create → name Kanutus, paste the URL → Connect.

You sign in with your Kanutus account and choose the organization and permissions on the Kanutus consent screen. Disconnect any time in Kanutus → Developers → Connected apps.

Claude Desktop, Claude Code, Cursor or your agent (local, API key)

Create an API key in Kanutus → Developers (use a key just for the assistant, with a spend limit), then:

{
  "mcpServers": {
    "kanutus": {
      "command": "npx",
      "args": ["-y", "@kanutus/mcp"],
      "env": { "KANUTUS_API_KEY": "kt_live_..." }
    }
  }
}

kt_test_... keys use the simulator: nothing uses minutes, test rooms have robot participants and magic phone numbers (+1 555 0100 answers, +1 555 0101 busy, +1 555 0102 voicemail, +1 555 0199 blocked).

Optional env: KANUTUS_API_URL (default https://api.kanutus.com).

Tools

Safety:

• Actions that spend minutes or reach other people ask you first (MCP elicitation). Clients without elicitation must pass confirm: true, which the assistant should only do after asking you.
• Anything said or written by other people (transcripts, room titles, names) comes back inside untrusted_content: it is data, never instructions.
• Create requests carry an idempotency key, so an assistant retrying a call never creates something twice.
• Live AI agents in rooms, voice cloning and phone calls are rolling out; with a live key they may answer “coming soon”. Try them with a kt_test_ key (simulator). In a room the agent always shows with the “AI agent” badge.

Changelog

• 0.3.7 (2026-10-28): create_room and schedule_meeting accept bill_mode (host: the host pays for every participant, default; each: every Kanutus account pays its own share) and guest_voice_clone (default true; false hides the opt-in that lets a guest without an account have their voice cloned for the meeting; the copy is deleted 3 days after last use). Nothing else changes.
• 0.3.6 (2026-10-26): AI avatars are coming soon. list_avatars returns an empty list with status: "coming_soon" and join_room accepts avatar_id but the agent joins voice only (the answer says so). Nothing else changes.
• 0.3.5 (2026-10-23): turn taking. listen and say return speaking (who is talking right now) and listen wakes up when a person starts or stops talking: wait until nobody is talking before say. The room never plays the agent over a person or over its own previous sentence, and a person who talks over the agent stops its voice (utterance.interrupted with spoken_text/unspoken_text). say takes an optional cursor and then returns what people said meanwhile. New stop_speaking tool.
• 0.3.4 (2026-10-22): say returns as soon as the speech is queued; call listen right away and keep calling it while you speak (people keep talking and their words only reach you on listen). Joining never waits for the engines to warm up: the agent enters at once (badge, no voice yet), uses the chat and watches engines/stage/eta_s.
• 0.3.3 (2026-10-22): the agent reads and writes the room chat. join_room returns chat_history (last 50 messages), listen returns a chat.message event per new message (text + translation to the hear language, null while the translator warms up) and the new send_chat posts in the chat as the agent (“AI agent” badge, translated for each reader; free; 2000 characters, one per second).
• 0.3.2 (2026-10-21): the agent’s introduction is automatic and mandatory. When it joins (and again to each person who joins later) the room hears and reads, translated, “Hello, I am , an AI agent.” and, with see_participants, “Depending on the situation, I may ask you to share your camera or your screen.” Free, no opt-out; listen returns agent.introduced. join_room now explains that evaluating people from what the agent sees (an interview, for example) needs their explicit consent beforehand and a human review (rules vary by country: EU, Illinois, New York, LGPD).
• 0.3.1 (2026-10-21): list_avatars (published avatar catalog) and avatar_id in join_room: the agent appears on camera with a Kanutus avatar animated by its voice; listen reports face: starting | live | unavailable. Agent vision: see_participants + snapshot_every_s in join_room (5, 10, 20, 40, 60, 90, 120, 500 or 900 s), snapshots of the participants’ cameras and screens as image blocks in listen, and the new look tool; the room shows everyone a permanent notice. Nothing is stored; no charge for snapshots in this version.
• 0.3.0: 27 languages (European Portuguese), numbers to call from (list_numbers, from in calls).

Develop

npm install
npm test                               # SDK client against the server (in memory and Streamable HTTP + OAuth)
npm run build                          # dist/ for npm
KANUTUS_API_KEY=kt_test_... npm run stdio
npm run dev                            # remote mode on http://localhost:3011/mcp (reads .env)

Remote mode env: KT_MCP_URL (this resource, e.g. https://mcp.kanutus.com/mcp), KT_API_URL (https://api.kanutus.com), KT_MCP_CLIENT_ID (kanutus-mcp) and KT_MCP_CLIENT_SECRET (the same value set in the API). How it authorizes:

1. Without a token, POST /mcp answers 401 with WWW-Authenticate: Bearer resource_metadata=".../.well-known/oauth-protected-resource/mcp".
2. That metadata names the Kanutus authorization server (https://api.kanutus.com, RFC 8414 metadata, PKCE S256, Dynamic Client Registration). The person consents at https://app.kanutus.com/oauth/authorize.
3. The client sends the token (audience = this MCP). The server checks it against the API JWKS and exchanges it (RFC 8693) for an API token with the same permissions. The user’s token is never forwarded.

Deploy (Vercel)

Project kanutus-mcp, root = this folder (or the root of the kanutus/mcp repo). api/index.ts handles every route (vercel.json rewrite). Env: the four variables above with NODE_ENV=production. Domain mcp.kanutus.com.

Publish to npm

npm run build && npm publish --access public

The kanutus/mcp public repository is exported from the Kanutus monorepo by scripts/kanutus-publish.sh mcp.