Key takeaways
- ACP is no longer independently developed — it merged into A2A under the Linux Foundation on August 29, 2025, and the GitHub repo is archived
- The official docs now lead with "ACP is now part of A2A" and publish a migration guide; BeeAI Platform itself switched to A2A
- IBM kept a governance seat — Kate Blair (IBM Research) joined the A2A Technical Steering Committee alongside Google, Microsoft, AWS, and others
- Naming collision — "ACP" in developer conversation today usually means Zed's Agent Client Protocol, an unrelated, actively developed standard
FAQ
What is ACP?
ACP (Agent Communication Protocol) was an open protocol from IBM's BeeAI team enabling AI agents, applications, and humans to communicate using multimodal messages over standard REST endpoints. It is now part of A2A.
Is ACP still maintained?
No. ACP merged into A2A under the Linux Foundation in August 2025. The GitHub repository (i-am-bee/acp) is archived, the team wound down independent development, and the docs point users to an A2A migration guide.
How did ACP differ from A2A?
ACP used REST/HTTP conventions and conversation-oriented messaging; A2A uses JSON-RPC 2.0 and task lifecycle management. The teams judged the approaches close enough to merge, with ACP's assets and expertise folded into A2A.
Is ACP the same as Zed's Agent Client Protocol?
No. Zed's Agent Client Protocol (also abbreviated ACP) is an unrelated, actively developed standard for connecting coding agents to editors. Since IBM's ACP merged into A2A, the Zed protocol is what most developers now mean by "ACP."
Executive Summary
Status note (June 2026): ACP is no longer an independent protocol. On August 29, 2025, IBM and the Linux Foundation announced that ACP would merge into A2A (Agent2Agent), with the ACP team winding down development and contributing its technology and expertise to A2A. [1] The GitHub repository is archived (last push August 25, 2025), and the official docs now open with "ACP is now part of A2A under the Linux Foundation" plus a migration guide. [2] [3]
ACP (Agent Communication Protocol) was IBM BeeAI's open protocol for agent-to-agent and agent-to-human communication, built on standard REST conventions. Where A2A uses JSON-RPC and task lifecycle management, ACP took a simpler approach: agents communicated via multimodal messages over familiar HTTP endpoints. [4] The merger resolved that split — IBM's Kate Blair joined the A2A Technical Steering Committee alongside Google, Microsoft, AWS, Cisco, Salesforce, ServiceNow, and SAP, and the BeeAI Platform itself moved from ACP to A2A. [1]
Naming collision: "ACP" today most often refers to Zed's Agent Client Protocol — an unrelated, actively developed standard (3.4K stars, commits as of June 2026) for connecting coding agents to editors. Don't confuse the two. [5]
| Attribute | Value |
|---|---|
| Organization | IBM (BeeAI / i-am-bee), Linux Foundation |
| Launched | March 2025 |
| Status | Merged into A2A (Aug 2025); repo archived |
| License | Apache 2.0 |
| GitHub Stars | ~1,013 (frozen at archive) |
| SDKs | Python, TypeScript (unmaintained) |
Product Overview
ACP enabled agents to communicate using rich, multimodal messages — not just text strings. [3] The design below is preserved for historical reference; new work happens in A2A.
Key Capabilities
| Capability | Description |
|---|---|
| Agent Manifest | Discovery document describing capabilities, name, and metadata |
| Runs | Single agent execution with inputs, supporting sync, async, and streaming |
| Multimodal Messages | Text, code, files, media in a single message structure |
| Sessions | Stateful conversation context across multiple runs |
| Distributed Sessions | URI-based session continuity across server instances |
| Trajectory Metadata | Track multi-step reasoning and tool calls across message parts |
Protocol Design
| Aspect | Detail |
|---|---|
| Transport | REST/HTTP(S) |
| Wire Format | JSON (OpenAPI-specified) |
| Discovery | Agent Manifests |
| Communication | Sync, streaming (SSE), async |
| State | Session-based (optional) |
| HA Support | Redis/PostgreSQL backends |
Strengths
- REST simplicity — Standard HTTP patterns integrated into existing infrastructure without specialized clients
- Multimodal native — Rich message types beyond text from day one
- Trajectory metadata — Useful feature for tracing multi-step agent reasoning
- Graceful exit — Rather than fragmenting the ecosystem, the team merged into A2A with a published migration path [1]
- Ideas live on — IBM holds an A2A Technical Steering Committee seat, carrying ACP's expertise into the surviving standard [1]
Cautions
- Protocol is sunset — Independent development ended August 2025; the repo is archived with zero open issues and no further releases [2]
- SDKs unmaintained — Python and TypeScript SDKs are frozen; no security or compatibility updates
- Migration required — Existing ACP deployments should follow the official migration guide to A2A [3]
- Stale tutorials — The DeepLearning.AI course and much 2025 "MCP vs A2A vs ACP" content still describe ACP as a live alternative; treat it as historical [6]
- Acronym confusion — Searching "ACP" now mostly surfaces Zed's Agent Client Protocol, a different protocol entirely [5]
What Developers Say
Reaction to the merger was muted — more shrug than mourning:
"IBM announced in March 2025 its Agent Communication Protocol (ACP) but is now abandoning the ACP name and merging ACP efforts with Google's Agent2Agent (A2A) protocol at the Linux Foundation." — schwentkerr, Hacker News [7]
"That seems odd. Even with an A2A protocol, don't you still need to standardize a client 'surface' or 'API' or whatever, so agents can describe IDE actions they want to trigger in the expected terms over that protocol?" — derefr, Hacker News [7]
From the principals: Kate Blair (IBM Research) framed the merger as consolidation — "By bringing the assets and expertise behind ACP into A2A, we can build a single, more powerful standard for how AI agents communicate and collaborate." [1]
Pricing & Licensing
Not applicable — ACP was a free, Apache 2.0 open-source protocol and remains freely available in archived form. [2]
Competitive Positioning (Historical)
| Protocol | ACP Differentiator (pre-merger) |
|---|---|
| A2A | ACP was REST-native and simpler; A2A won on adoption (24K+ stars, June 2026) and absorbed ACP [1] |
| ANP | ACP was centralized/enterprise; ANP is decentralized with DID-based identity |
| MCP | MCP is agent-to-tool; ACP was agent-to-agent. Complementary |
| Agent Client Protocol (Zed) | Unrelated despite the shared acronym — agent-to-editor, actively developed [5] |
Ideal Customer Profile
Best fit:
- No one, for new builds — adopt A2A directly
- Historical reference for protocol designers studying REST-based agent communication
Poor fit:
- Any team starting a new agent interoperability project in 2026
- Existing ACP deployments that haven't migrated — the migration guide to A2A is the supported path [3]
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | N/A — absorbed into A2A (Linux Foundation) |
| Market Position | Sunset — merged into the winning standard |
| Innovation Pace | Ended — repo archived August 2025 [2] |
| Community/Ecosystem | Dissolved into A2A's (24K+ stars, 8-company TSC) [1] |
| Long-term Outlook | Resolved — the February 2026 prediction of A2A convergence came true (it had in fact already happened) |
Bottom Line
ACP was a well-designed, REST-native protocol that was simpler to adopt than A2A — but the question of which would win has been answered. ACP merged into A2A under the Linux Foundation in August 2025, its repo is archived, and even the BeeAI Platform that ACP was built for now runs on A2A. [1]
Recommended for: Historical study only. The multimodal messaging and trajectory metadata ideas are worth reading as design input for A2A-based systems.
Not recommended for: Any new project. Build on A2A (agent-to-agent) and MCP (agent-to-tool); if you saw "ACP" mentioned in an editor/coding-agent context, that's Zed's separate Agent Client Protocol. [5]
Outlook: Closed chapter. ACP's legacy is its contribution to A2A — IBM retains a Technical Steering Committee seat, and the consolidation left the agent-protocol landscape clearer than it found it. [1]
Research by Ry Walker Research • methodology