io.github.Oortonaut/mcacp
编码与调试by oortonaut
连接 MCP 与 ACP 的桥接工具,让任意 MCP client 都能驱动 ACP coding agents。
什么是 io.github.Oortonaut/mcacp?
连接 MCP 与 ACP 的桥接工具,让任意 MCP client 都能驱动 ACP coding agents。
README
MCACP
Bridge any MCP client to any ACP coding agent.
MCACP is an MCP server that speaks the Agent Client Protocol on behalf of your MCP host. It lets Claude Code, Zed, VS Code, or any MCP-compatible client spawn, manage, and interact with ACP coding agents — turning a single chat window into a multi-agent control plane.
MCP Client (Claude Code, Zed, VS Code, ...)
|
| MCP (stdio)
v
MCACP
|
| ACP (stdio, per agent)
v
Claude Code · Gemini CLI · Codex · Auggie · Copilot · ...
What can you do with this?
Run any ACP agent from any MCP client. You don't need a dedicated IDE plugin for each agent. Install Gemini from Claude Code. Run Codex from Zed. Mix and match.
Compare agents side by side. Give the same task to Claude Code ACP and Gemini CLI in parallel sessions and compare the results.
Build agent-driven pipelines. One agent writes code, another reviews it, a third runs the tests. MCACP handles the lifecycle and permission flow for all of them.
Control permissions interactively. On hosts that support MCP elicitation (Zed), permission requests pop up as interactive forms. On hosts that don't (Claude Code), they fall back to operator mode where the host agent decides.
Resume where you left off. Sessions persist to disk. Shut down, restart, load_session — the agent picks up where it stopped.
Discover and install agents from chat. Browse the ACP registry, install agents, check for upgrades — without leaving your conversation.
Quick start
Install
npm install -g mcacp
Or run directly:
npx mcacp
Add to your MCP client
Claude Code (~/.claude/settings.json):
{
"mcpServers": {
"mcacp": {
"command": "npx",
"args": ["-y", "mcacp"]
}
}
}
Zed (settings.json):
{
"context_servers": {
"mcacp": {
"command": { "path": "npx", "args": ["-y", "mcacp"] }
}
}
}
Then from chat:
"Search for available agents" → registry_search
"Install the Gemini agent" → agent_install
"Initialize it and start a session" → initialize + new_session
"Ask it to refactor auth" → prompt_start + prompt
MCP tools
MCACP exposes 22 tools organized into five groups.
Registry
| Tool | Description |
|---|---|
list_installed_agents | List locally installed agents |
registry_search | Search the ACP agent registry |
agent_install | Install an agent from the registry |
agent_uninstall | Remove an installed agent |
agent_check_upgrades | Check for available upgrades |
discover_agents | Scan editor configs (Zed, JetBrains) for agents |
reload_config | Reload mcacp.json from disk |
Lifecycle
| Tool | Description |
|---|---|
initialize | Spawn an agent process and perform the ACP handshake |
shutdown | Gracefully shut down an agent |
Sessions
| Tool | Description |
|---|---|
new_session | Create a session with working directory and permission policy |
load_session | Resume a persisted session |
list_sessions | List stored sessions for an agent |
close_session | Close a session (preserved for later resume) |
Interaction
| Tool | Description |
|---|---|
prompt_start | Send a prompt (returns immediately) |
prompt | Block until events are available |
prompt_events | Non-blocking poll for events |
grant_permission | Resolve a pending permission request (operator mode) |
request_permission | Send an MCP elicitation to the host for a permission decision |
cancel | Cancel an in-progress prompt |
set_mode | Switch agent operating mode |
Status
| Tool | Description |
|---|---|
list_running_agents | List all agents with status and sessions |
get_agent_status | Detailed status of a running agent |
set_agent_status | Set status text for an agent |
Permission policies
Each session can use a different policy:
| Policy | Behavior |
|---|---|
allow_all | Auto-approve all agent actions |
deny_all | Block all tool use |
elicit | Forward as MCP elicitations to the host; falls back to operator if unsupported |
operator | Queue as events — the host agent decides via grant_permission |
Operator flow
Orchestrator MCACP ACP Agent
| | |
|-- prompt(sessionId, ...) ->| |
| |-- session/prompt --------->|
| |<- request_permission ------|
|<- {pendingPermission: ...} | |
| | |
|-- grant_permission ------->| |
| {toolCallId, optionId} |-- permission outcome ----->|
| |<- prompt result -----------|
|<- {stopReason, updates} | |
Configuration
MCACP looks for mcacp.json in these locations (first found wins):
--configflag orMCACP_CONFIGenv var.mcacp/mcacp.jsonin the current directory~/.config/mcacp/mcacp.json
{
"registries": ["https://cdn.agentclientprotocol.com/registry/v1/latest/registry.json"],
"defaultPermissionPolicy": "elicit",
"agent_servers": {
"my-custom-agent": {
"command": "node",
"args": ["my-agent.js"]
}
}
}
See docs/configuration.md for the full schema.
Building from source
git clone https://github.com/Oortonaut/mcacp.git
cd mcacp
npm install
npm run build
node dist/index.js
License
常见问题
io.github.Oortonaut/mcacp 是什么?
连接 MCP 与 ACP 的桥接工具,让任意 MCP client 都能驱动 ACP coding agents。
相关 Skills
网页构建器
by anthropics
面向复杂 claude.ai HTML artifact 开发,快速初始化 React + Tailwind CSS + shadcn/ui 项目并打包为单文件 HTML,适合需要状态管理、路由或多组件交互的页面。
✎ 在 claude.ai 里做复杂网页 Artifact 很省心,多组件、状态和路由都能顺手搭起来,React、Tailwind 与 shadcn/ui 组合效率高、成品也更精致。
前端设计
by anthropics
面向组件、页面、海报和 Web 应用开发,按鲜明视觉方向生成可直接落地的前端代码与高质感 UI,适合做 landing page、Dashboard 或美化现有界面,避开千篇一律的 AI 审美。
✎ 想把页面做得既能上线又有设计感,就用前端设计:组件到整站都能产出,难得的是能避开千篇一律的 AI 味。
网页应用测试
by anthropics
用 Playwright 为本地 Web 应用编写自动化测试,支持启动开发服务器、校验前端交互、排查 UI 异常、抓取截图与浏览器日志,适合调试动态页面和回归验证。
✎ 借助 Playwright 一站式验证本地 Web 应用前端功能,调 UI 时还能同步查看日志和截图,定位问题更快。
相关 MCP Server
GitHub
编辑精选by GitHub
GitHub 是 MCP 官方参考服务器,让 Claude 直接读写你的代码仓库和 Issues。
✎ 这个参考服务器解决了开发者想让 AI 安全访问 GitHub 数据的问题,适合需要自动化代码审查或 Issue 管理的团队。但注意它只是参考实现,生产环境得自己加固安全。
Context7 文档查询
编辑精选by Context7
Context7 是实时拉取最新文档和代码示例的智能助手,让你告别过时资料。
✎ 它能解决开发者查找文档时信息滞后的问题,特别适合快速上手新库或跟进更新。不过,依赖外部源可能导致偶尔的数据延迟,建议结合官方文档使用。
by tldraw
tldraw 是让 AI 助手直接在无限画布上绘图和协作的 MCP 服务器。
✎ 这解决了 AI 只能输出文本、无法视觉化协作的痛点——想象让 Claude 帮你画流程图或白板讨论。最适合需要快速原型设计或头脑风暴的开发者。不过,目前它只是个基础连接器,你得自己搭建画布应用才能发挥全部潜力。