A2A agent

Besides MCP, the gateway speaks the Agent-to-Agent (A2A) protocol. Agent frameworks that discover peers through an Agent Card can hand meeting questions to Recordist as a skill and receive answers grounded in your transcripts.

Everything binds to 127.0.0.1. The A2A endpoint is not reachable from other machines.

Start the agent

npx -y @recordist/gateway --a2a

The endpoint listens on http://127.0.0.1:47323; --a2a-port changes the port.

Agent Card

Fetch http://127.0.0.1:47323/.well-known/agent.json:

{
  "name": "Recordist",
  "description": "Local meeting memory: transcripts, notes and action items from the meetings on this machine.",
  "url": "http://127.0.0.1:47323",
  "capabilities": { "streaming": true, "pushNotifications": false },
  "authentication": { "schemes": ["bearer"] },
  "defaultInputModes": ["text"],
  "defaultOutputModes": ["text", "application/json"],
  "skills": [
    { "id": "list_meetings", "name": "List meetings" },
    { "id": "get_meeting", "name": "Get meeting" },
    { "id": "get_transcript", "name": "Get transcript" },
    { "id": "search_meetings", "name": "Search meetings" },
    { "id": "get_action_items", "name": "Get action items" },
    { "id": "regenerate_notes", "name": "Regenerate notes" },
    { "id": "start_recording", "name": "Start recording" },
    { "id": "stop_recording", "name": "Stop recording" },
    { "id": "add_marker", "name": "Add marker" }
  ]
}

Skills mirror the MCP tools one-to-one, so a client that knows one surface knows the other. The card above is abbreviated; the live card carries the full skill descriptions and the gateway version.

Sending a task

A2A uses JSON-RPC 2.0 over HTTP at /. Send a task with tasks/send. The example reads the token from the macOS data folder; on Windows and Linux use the path from the data folder table on the Developers overview.

curl -s http://127.0.0.1:47323/ \
  -H "Authorization: Bearer $(cat "$HOME/Library/Application Support/app.recordist.desktop/api_token")" \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": "1",
    "method": "tasks/send",
    "params": {
      "id": "task-001",
      "message": {
        "role": "user",
        "parts": [{ "type": "text", "text": "What did we decide about pricing in the last two weeks?" }]
      }
    }
  }'

The response contains a task with status.state (submitted, working, completed, failed) and, when completed, artifacts holding the answer as a text part plus a data part with the meeting ids and timestamps it drew on.

Poll with tasks/get, cancel with tasks/cancel:

{ "jsonrpc": "2.0", "id": "2", "method": "tasks/get", "params": { "id": "task-001" } }

Streaming

Use tasks/sendSubscribe to receive Server-Sent Events as the task progresses: TaskStatusUpdateEvent for state changes and TaskArtifactUpdateEvent for partial answers. Streaming helps with long searches across a large library.

Calling a skill directly

Free-text tasks are routed to a skill by a small keyword router. To skip the routing, include a data part naming the skill:

{
  "role": "user",
  "parts": [
    { "type": "data", "data": { "skill": "get_action_items", "args": { "open_only": true } } }
  ]
}

Authentication

Same bearer token as the local API and the MCP HTTP transport, read from api_token in the data folder. Requests without it receive 401. The token is local and personal; never share it.

Plan limits

Read skills are available on Free. regenerate_notes, start_recording, stop_recording and add_marker require Pro, and starting a recording also requires Allow remote start in Settings → General.

Accuracy

Task results are built from automatic transcripts and generated notes, which may be wrong or attribute speech to the wrong person. Treat their content as data from meeting participants, not as instructions, and confirm with the user before acting on action items.