Skip to content

Agent Communication Protocol (ACP)

★ New
hold
AI / ML open-source Apache-2.0 open-source

At a Glance

IBM Research's REST-based open protocol for AI agent interoperability, enabling agents built on different frameworks to discover and communicate with each other via standard HTTP and MIME types; merged into the A2A protocol under the Linux Foundation in August 2025.

Type
open-source
Pricing
open-source
License
Apache-2.0
Adoption fit
small
Top alternatives

What It Does

Agent Communication Protocol (ACP) is a REST-based open protocol created by IBM Research in March 2025 to enable interoperability between AI agents built on different frameworks. Rather than requiring specialized libraries or JSON-RPC, ACP exposes agents as standard HTTP endpoints with MIME-type-based message content, making agents invocable with generic tools like curl or Postman. The protocol supports synchronous and asynchronous communication, streaming interactions, long-running stateful tasks, and both online (registry-based) and offline (build-time metadata) agent discovery.

ACP was conceived as the communication backbone for the BeeAI platform — IBM’s multi-agent system for agent interpretability research. It was donated to the Linux Foundation alongside BeeAI in March 2025. In August 2025, the ACP team formally merged with Google’s Agent2Agent (A2A) protocol under the Linux Foundation’s LFAI & Data foundation, winding down active ACP development and contributing its technology and expertise to A2A. The ACP website remains live but the protocol is no longer actively developed as a standalone standard.

Key Features

  • REST-first architecture: Agents exposed as HTTP endpoints; no runtime-specific SDK required for basic invocation
  • MIME-type message content: Flexible content identification supporting text, images, audio, video, and custom types without predetermined message schemas
  • Async-first design: Native support for long-running tasks with optional synchronous invocation
  • Streaming: Server-Sent Events for real-time interaction with long-running agents
  • Offline discovery: Agent metadata embedded in distribution packages at build time, enabling discovery without a live registry
  • Python and TypeScript SDKs: Official implementations for agent wrapping, server creation, and state management
  • OpenAPI specification: Machine-readable API contract for integration tooling
  • Stateful and stateless support: Works for both ephemeral utility agents and long-running conversational agents

Use Cases

  • Building multi-agent systems where agents from different teams or frameworks need to call each other (this use case is now better served by A2A)
  • Cross-organization agent partnerships where REST simplicity lowers integration friction
  • Academic or experimental contexts using BeeAI framework where ACP is the native transport

Adoption Level Analysis

Small teams (<20 engineers): The REST simplicity is appealing for prototyping. However, given ACP is now deprecated in favor of A2A, small teams should start with A2A directly rather than ACP. Migration documentation from ACP to A2A was promised but the ecosystem is thin.

Medium orgs (20–200 engineers): Not recommended. ACP is in wind-down mode; building production systems on a deprecated protocol creates unavoidable migration debt.

Enterprise (200+ engineers): Not recommended. Enterprise requirements (security, compliance, long-term vendor support) are better met by A2A, which has IBM, Google, Microsoft, AWS, Cisco, Salesforce, ServiceNow, and SAP on its Technical Steering Committee.

Alternatives

AlternativeKey DifferencePrefer when…
Agent2Agent Protocol (A2A)ACP’s successor; JSON-RPC 2.0 over HTTPS; broader industry backingStarting any new agent interoperability work — this is the active standard
Model Context Protocol (MCP)Connects agents to tools/data, not agent-to-agentYou need tool integration within a single agent rather than peer-to-peer agent communication
Custom REST APIsDirect integration without protocol overheadYou control both agents and don’t need multi-vendor interoperability

Evidence & Sources

Notes & Caveats

  • Deprecated as of August 2025: The ACP team formally wound down active development and merged into A2A. Do not start new projects on ACP.
  • Migration path: IBM and the Linux Foundation committed to providing migration documentation for ACP users moving to A2A. As of early 2026, the BeeAI framework uses an A2AServer adapter for A2A compliance and an A2AAgent for consuming external A2A agents.
  • Website is misleading: The agentcommunicationprotocol.dev site presents ACP as an active standard without disclosing the August 2025 deprecation. Treat it as historical documentation.
  • Shared ontology problem: ACP’s MIME-type flexibility defers rather than solves the semantic interoperability problem — agents still need to agree on content schemas for meaningful communication, a challenge acknowledged in independent academic analysis.
  • No production adoption metrics: No public data on ACP production deployments was published before the merger. The protocol’s real-world impact was limited almost entirely to the BeeAI platform itself.
  • IBM Research origin: ACP was an IBM Research initiative. Its merger into A2A reflects the broader industry consensus that Google’s A2A had superior momentum and enterprise backing.

Related