A2A protocol

The Agent-to-Agent (A2A) protocol lets AI agents communicate with your deployed functions using a standard JSON-RPC interface. Every organization with deployed MCP tools automatically gets an A2A endpoint — no extra configuration needed.

Agent Card

Each organization publishes a discoverable Agent Card at:

GET https://{tenant}.fold.run/.well-known/agent.json

The card is auto-generated from your deployed MCP tools:

{
  "name": "fold-my-tenant",
  "description": "AI agent powered by fold.run with 3 tool(s)",
  "url": "https://my-tenant.fold.run",
  "version": "0.1.0",
  "capabilities": {
    "streaming": true,
    "pushNotifications": false
  },
  "skills": [
    {
      "id": "weather-lookup",
      "name": "weather-lookup",
      "description": "Look up current weather for a city",
      "inputModes": ["application/json"],
      "outputModes": ["application/json"]
    }
  ],
  "defaultInputModes": ["application/json"],
  "defaultOutputModes": ["application/json"]
}

AI agents use this card to discover what skills your organization offers.

Sending tasks

Send a task to your A2A endpoint:

curl -X POST https://my-tenant.fold.run/_a2a/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": "1",
    "method": "tasks/send",
    "params": {
      "skill": "weather-lookup",
      "message": { "city": "San Francisco" }
    }
  }'

Response:

{
  "jsonrpc": "2.0",
  "id": "1",
  "result": {
    "id": "task_abc123",
    "status": { "state": "completed" },
    "artifacts": [
      {
        "name": "weather-lookup-result",
        "parts": [{ "type": "text", "text": "{\"temp\":62,\"condition\":\"foggy\"}" }]
      }
    ]
  }
}

Checking task status

Retrieve the status of a previously submitted task:

curl -X POST https://my-tenant.fold.run/_a2a/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": "2",
    "method": "tasks/get",
    "params": { "id": "task_abc123" }
  }'

Canceling tasks

Cancel a task that is still in the working state:

curl -X POST https://my-tenant.fold.run/_a2a/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": "3",
    "method": "tasks/cancel",
    "params": { "id": "task_abc123" }
  }'

Only tasks with status submitted or working can be canceled.

Task lifecycle

Tasks move through these states:

State	Description
submitted	Task received, queued for execution
working	Currently executing
completed	Finished successfully (artifacts available)
failed	Execution errored
canceled	Canceled by caller

Tasks are retained for 24 hours and then automatically cleaned up.

JSON-RPC methods

Method	Description
tasks/send	Execute a skill with input data
tasks/sendSubscribe	Execute a skill and stream status/artifact events via SSE
tasks/get	Retrieve task status and artifacts
tasks/cancel	Cancel a running task

Error codes

Standard JSON-RPC error codes apply:

Code	Meaning
-32600	Invalid request
-32601	Method not found
-32602	Invalid params (e.g. unknown skill)
-32603	Internal error

Streaming tasks

Use tasks/sendSubscribe to receive real-time SSE events as a task executes. See the Streaming guide for details and examples.

Pipelines as skills

Active pipelines are automatically exposed as A2A skills with pipeline: prefix. For example, a pipeline named summarize-and-translate appears as skill pipeline:summarize-and-translate in the Agent Card.

Guardrails

If guardrail policies are configured, inputs are checked before dispatch and outputs are checked after completion. A block policy returns a JSON-RPC error instead of executing the skill.

How it relates to MCP

MCP and A2A are complementary:

• MCP is for tool-use — an AI agent calls your function as a tool within its reasoning loop.
• A2A is for agent-to-agent delegation — one agent asks another agent to complete a task.

Both protocols are auto-configured from the same deployed functions and pipelines. Deploy once, and your code is reachable via MCP (/mcp), A2A (/_a2a/), and plain HTTP.