A2A agent

Vidmoat answers the Agent2Agent protocol with one skill: edit a project from a brief. A task is the same server-side run of the editing agent that the Vidmoat editor starts, with the same gates, the same per-step credit charge and the same result on the timeline. Use MCP when your agent wants the individual tools; use A2A when it wants to delegate the whole edit.

What it does not do#

• No streaming and no push notifications. Both are false on the card. SendMessage returns the task at once, submitted; poll GetTask until it is completed, failed or canceled.
• No file parts. Upload media to the project first (MCP upload_media or POST /v1/media) and describe the edit in a text part. A file part is refused with -32005, not ignored.
• No project creation. Create one with MCP create_project or POST /v1/projects and pass its id.
• No follow-up on a finished task. Every task ends in a terminal state. Send a new message with referenceTaskIds naming the earlier task (up to three); the agent gets what was asked and what it said.
• No batches and no ListTasks. Keep the task ids you are given.
• Test keys cannot start runs, because runs spend real credits.
• Server-side runs need a plan that includes them (Creator and above) with server-side runs turned on for the account; otherwise SendMessage answers -32000 with data.code not_entitled or not_opted_in.

Send a brief#

The project id goes in message.metadata.projectId, in a data part {"projectId": "..."}, or as the contextId. A context is one project: the task's contextId is always the project id.

curl https://api.vidmoat.com/api/a2a \
  -H "Authorization: Bearer $VIDMOAT_API_KEY" \
  -H "A2A-Version: 1.0" \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0", "id": 1, "method": "SendMessage",
    "params": { "message": {
      "messageId": "m-1", "role": "ROLE_USER",
      "parts": [{ "text": "Cut the silences and add bold captions." }],
      "metadata": { "projectId": "prj_123" }
    } }
  }'

{ "jsonrpc": "2.0", "id": 1, "result": { "task": {
  "id": "5d0c…", "contextId": "prj_123",
  "status": { "state": "TASK_STATE_SUBMITTED", "timestamp": "2026-10-26T12:00:00.000Z" },
  "metadata": { "projectId": "prj_123", "editorUrl": "https://www.vidmoat.com/editor?project=prj_123", "applied": 0, "attempted": 0 }
} } }

Poll it#

{ "jsonrpc": "2.0", "id": 2, "method": "GetTask", "params": { "id": "5d0c…", "historyLength": 5 } }

A completed task carries the agent's closing message in status.message, its messages in history, and one artifact, Edited project, whose data part names the project, the editor link, the commands applied and attempted, and the credits left when the run last reported them. A failed task carries the reason in status.message.

CancelTask asks the run to stop at its next step boundary, so a half-applied batch never reaches the document. The task stays working (with metadata.cancelRequested) until it does, then reads canceled.

Errors#

Google models are not used for runs started here, as for MCP, the REST API and the chat bots.

Skills for your own agent#

The editing craft the agent runs on is published as Agent Skills: the index is at https://vidmoat.com/.well-known/agent-skills/index.json (v0.2.0), each SKILL.md is served beside it with its sha256 digest in the index, and the source is the public repository vidmoat/skills.