acp

ACP — Agent Communication Protocol

Overview

ACP is an open protocol (originally from IBM / BeeAI, now contributed to the Linux Foundation alongside A2A) for communication between AI agents, applications, and humans. It uses plain REST endpoints — no JSON-RPC, no specialized transport — making it usable with standard HTTP tools like curl, Postman, or any HTTP client.

Status: The standalone ACP repository is archived. ACP's design has been merged into the A2A project under the Linux Foundation. New projects should evaluate A2A, but existing ACP deployments and the SDK remain functional.

Core Concepts

Agent Manifest

Every ACP agent exposes a manifest describing its capabilities — without revealing implementation details:

{
  "name": "research-agent",
  "description": "Finds and summarizes information on any topic",
  "metadata": {
    "capabilities": ["streaming", "sessions"]
  }
}

Messages and Parts

Agents exchange Messages composed of Parts. Each part carries a MIME type, enabling multimodal payloads (text, images, audio, files) without protocol changes:

{
  "role": "user",
  "parts": [
    {
      "content": "Summarize this document",
      "content_type": "text/plain",
      "content_encoding": "plain"
    },
    {
      "content": "<base64-data>",
      "content_type": "application/pdf",
      "content_encoding": "base64",
      "name": "report.pdf"
    }
  ]
}

Part Metadata

Parts can carry structured metadata for observability and attribution:

• TrajectoryMetadata — exposes reasoning steps and tool calls
• CitationMetadata — attributes content to source documents

REST API

Endpoints

Method	Path	Description
GET	/agents	List available agents with their manifests
POST	/runs	Execute an agent (sync, async, or streaming)

Run Lifecycle

pending → running → [awaiting] → completed / error

State	Meaning
pending	Run queued, not yet started
running	Agent is executing
awaiting	Agent paused, requesting external input
completed	Finished successfully
error	Execution failed