ACP adapter for the Claude Agent SDK

Use Claude Agent SDK from ACP-compatible clients!

This tool implements an ACP agent by using the official Claude Agent SDK, supporting:

• Context @-mentions
• Images
• Tool calls (with permission requests)
• Compact file changes through the negotiated AIR diff patch extension
• Following
• Edit review
• TODO lists
• Nested subagent transcripts
• Interactive (and background) terminals
• Custom Slash commands
• Client MCP servers
• An Ultracode session config option, which turns on Claude Code's dynamic workflows for every task. It appears when dynamic workflows are enabled and the model supports ultracode (any model with xhigh effort)
• /mcp in the chat: the MCP server status as a list. The adapter runs /mcp reconnect, /mcp enable, and /mcp disable through the SDK control API, because Claude Code refuses them in SDK mode. A reconnect of an ACP server that needs authentication starts MCP OAuth through URL elicitation
• Session-scoped long-running goals for AIR through the goal extension under _meta.jetbrains.air.goal
• Background task lifecycle and per-task stop control through the opt-in async tasks extension, including the provider-neutral clientCapabilities._meta["async-tasks"]: true negotiation
• Structured errors, recovery, and warnings through the opt-in session failure extension
• Concrete model and effort defaults through the opt-in recommended config value extension
• Tool permission presentation, editable choices, and durable effects through the permission extension
• One fact per field in every tool call report, as the ACP tool call contract defines
• All AIR extensions, capabilities, and _meta keys: AIR extensions

Learn more about the Agent Client Protocol.

To try changes that have landed on main but are not released yet, install from the preview channel — every push to main publishes one. See docs/RELEASES.md.

npm install @agentclientprotocol/claude-agent-acp@preview

Standalone executables

Each release attaches archives for Linux (glibc and musl), macOS and Windows, on x64 and arm64, to its GitHub release. They need no Node.js install. Each holds two files: claude-agent-acp, and the Claude CLI (claude) it runs. Keep them in the same directory; a symlink to claude-agent-acp elsewhere works. To use a Claude CLI you already have instead, set CLAUDE_CODE_EXECUTABLE. On Alpine, install libgcc and libstdc++ first.

To build one for your machine (needs Bun) and check it:

npm run build:binary
npm run smoke:binary            # add `-- --full` to drive a real session; needs a Claude login

Subagent sessions

A client that declares clientCapabilities.subagents, other than JetBrains AIR, gets each Agent or Task subagent as its own child session, following the subagents RFD:

• subagent_update on the parent announces the child before any of its traffic, with the parent's title and description for it, and reports its work state: running, requires_action while one of its requests is open (and running again when the last one ends, unless a cancel withdrew it: the child is stopping then), and idle when its work ends, with end_turn or cancelled when the SDK gives the reason. A failure is idle with the error stop reason and the SDK's error text as its error. The child keeps its session for every delegation to it.
• A child that runs outside its parent's working directory, such as an Agent call with isolation: "worktree", gets its cwd on subagent_update before its first tool call. The adapter takes it from the cwd of the child's own hooks: its SubagentStart hook runs before its first model request. It sends cwd: null when the child is back in its parent's directory. A child in its parent's directory gets no cwd, and no child gets additionalDirectories: Claude Code gives a subagent the session's additional directories, so it inherits them. A replay announces a child with the cwd that its transcript recorded, and clears it after the response when that directory no longer exists.
• Session messages replace the Agent or Task call that launches the child, which its parent is not shown:
  • The launch's prompt is a session_message in both transcripts: the parent's outgoing view and the child's incoming one. Both use the launch's tool use id, and each names the other with the draft RFD's recipientMessageId and senderMessageId.
  • A foreground child's answer, from the launch's result, is the child's session_message to the parent. In Claude Code's hand-back mode, the child's SubagentHandback call is its message to the parent instead, in both transcripts and linked the same way.
  • A launch whose child is never exposed is shown as a tool call after all, with its final status.
• The message of a SendMessage call that resumes the child is a session_message in the child's transcript. The SendMessage call stays a tool call of the session that made it.
• session/cancel on the parent interrupts Claude Code, which stops the foreground children that the turn waits on. Background children keep running, as the RFD prefers; a client stops one with its own cancel. Each stopped child reports idle with cancelled from Claude Code's own events, and its open requests are withdrawn. The adapter ends no child at a cancel itself, so a child that keeps running keeps its open requests and its open tool calls. To spare background children, the adapter declares the SDK option perTaskStopAffordance, which a client can turn off in _meta.claudeCode.options. That option spares background workflows too, which a client cannot stop, so the adapter stops them after the interrupt, as Claude Code did without it. Background commands keep running after a cancel either way. A tool call of a stopped child that Claude Code reports no result for, such as the command it was running, ends in the child's session before the cancelled prompt resolves. When Claude Code stops a background child after that, or while no prompt runs, the tool call ends right after the child reports idle. A child that still runs when Claude Code exits is idle with no stop reason.
• A child whose parent is not yet known when it ends is not exposed, since the RFD forbids guessing its parent.
• A replay restores each child under its live session id, from its own transcript, with the same session messages as live, and its end is an idle before the response. A launch without Claude Code's metadata file for the child stays an ordinary tool call of the parent, and a notice, for a client that takes notices, says that subagent history is missing.
• After a load or resume, with or without replay, a SendMessage that resumes an agent from before it runs that agent in its session from before. Before the agent's traffic, each of its ancestors that the client was not told about in this connection is announced, parent before child and with no work state, as the RFD requires. An agent whose launch the session's history does not record stays unexposed.
• Each running child is announced with the cancel capability. session/cancel with the child's session id stops that child's current work, through the SDK's Query.stopTask, and not its parent's: Claude Code also stops the foreground children it waits on, and the parent's turn goes on. Each stopped child reports idle with cancelled from Claude Code's own events, and its requests that the stop withdraws report no running first. A replayed child, which no longer runs, gets no capabilities. A client cannot prompt or otherwise change a child.

An ACP v2 client gets the same without declaring anything, since subagents are part of v2's baseline. Each child also reports its work state as its own state_update, right after the parent's copy.

JetBrains AIR keeps the earlier draft of the RFD (subagent_spawned, subagent_state_update), which it enables with nativeSubagentSessions in _meta.jetbrains.air.capabilities. See AIR extensions.

Without either signal, Agent and Task stay ordinary tool calls and child interactions stay on the root session. Clients that use the historical _meta["subagent-transcript"] capability or the forwardSubagentText session option retain the flattened child transcript.

Experimental ACP v2

The adapter also serves the draft ACP v2: a client that asks for protocol version 2 or later in initialize gets v2, and every other client gets v1. The draft can still change, and this surface changes with it.

CLAUDE_AGENT_ACP_PROTOCOL limits the adapter to one version. With v1, a client that asks for v2 is negotiated down to v1. With v2, the adapter rejects the initialize of a v1 client. auto, the default, serves both.

On v2 the adapter serves the session baseline (session/new, session/list, session/resume with replay, session/close, session/prompt, session/cancel), session/delete, session/fork, additional directories, config options (the mode is one), plans, notices, compaction, display terminals for commands, permission requests, elicitation, providers, auth, and subagents. Steering and the AIR extensions are v1 only for now.

Contribution Policy

This project does not require a Contributor License Agreement (CLA). Instead, contributions are accepted under the following terms:

By contributing to this project, you agree that your contributions will be licensed under the Apache License, Version 2.0. You affirm that you have the legal right to submit your work, that you are not including code you do not have rights to, and that you understand contributions are made without requiring a Contributor License Agreement (CLA).