ACP Support

ACP is a standardized communication protocol that decouples AI coding agents from editors, letting both sides evolve independently.

Source: packages/opencorvus/src/acp/. SDK: @agentclientprotocol/sdk@0.14.1.

opencorvus acp

opencorvus acp                         # current directory
opencorvus acp --cwd /path/to/project
OPENCORVUS_ENABLE_QUESTION_TOOL=1 opencorvus acp

Source: packages/opencorvus/src/cli/cmd/acp.ts

On startup:

1. Internal opencorvus serve on a random port
2. Creates an @opencorvus-ai/sdk client against that internal server
3. Opens a JSON-RPC channel over stdin/stdout via ndJsonStream from @agentclientprotocol/sdk
4. Instantiates ACP.Agent, waits for client messages

Transport: newline-delimited JSON over stdio — no external TCP port (src/cli/cmd/acp.ts).

Connecting Zed

~/.config/zed/settings.json:

{
  "agent_servers": {
    "OpenCorvus": {
      "command": "opencorvus",
      "args": ["acp"]
    }
  }
}

Zed spawns opencorvus acp on demand and communicates via stdin/stdout (src/acp/README.md).

Supported ACP methods

Initialization

Method	Description
initialize	Negotiates protocol version (1), advertises capabilities

Advertised (src/acp/agent.ts): loadSession: true; MCP http, sse; prompt embeddedContext, image; session fork, list, resume.

Session management

Method	Description
session/new	Creates new session
session/load	Loads session, replays history, restores model and mode
unstable_listSessions	Lists sessions newest first, 100/page cursor
unstable_forkSession	Forks a session, replays history
unstable_resumeSession	Resumes without replay
unstable_setSessionModel	Switches model, supports variants
setSessionMode	Switches primary agent

Prompting

Method	Description
session/prompt	Sends message; text / image / resource_link / resource; /command prefix → slash command

Real-time notifications

Via connection.sessionUpdate():

Update	When
agent_message_chunk	Streaming AI text (driven by message.part.delta)
agent_thought_chunk	Reasoning model thoughts
tool_call	Tool invocation start
tool_call_update	In progress / done / failed
plan	Todo list sync on todowrite completion
usage_update	Tokens + cost after each prompt
available_commands_update	Slash commands on session load

Authentication

Caution

authenticate is not implemented — throws "Authentication not implemented" (src/acp/agent.ts).

Permission handling

ACP forwards durable permission.asked events through connection.requestPermission(). Eligible choices map to allow_once, allow_task, allow_project, or deny on the canonical permission request.

For edit permissions, after client approval, ACP calls connection.writeTextFile() to push the patched content to the editor (src/acp/agent.ts).

MCP server support

session/new and session/load accept an mcpServers list:

• Local processes (command, args, env)
• Remote HTTP/SSE (url, headers)

ACP calls sdk.mcp.add() to register each server into the internal opencorvus instance.

vs opencorvus serve (REST)

Dimension	ACP	serve
Transport	stdin/stdout (ndjson)	HTTP (REST + SSE)
Clients	ACP-compatible editors	Browser / SDK / GH Action
Protocol	ACP v1 (JSON-RPC)	Custom REST
Streaming	session/update	SSE /event
External port	None	7878
Concurrency	Single client	Multi-client

vs MCP

Dimension	ACP	MCP
Role	Editor ↔ AI agent	LLM ↔ tool/data
Direction	Editor calls agent	LLM calls tools
Sessions	Yes (session/new/load)	No
Permission UI	Agent asks editor	N/A
OpenCorvus	ACP Server (editor connects)	MCP Client (connects to external MCP)

Current limitations

(src/acp/README.md)

• Authentication: authenticate throws immediately
• Session persistence: session/load replays messages but does not restore working-directory snapshot
• Terminal support: terminal capability in client.ts is a stub
• Session mode switching: setSessionMode works but lacks UI feedback