Claude Code MCP
Delegates coding tasks to the local Claude Code CLI via MCP, offering run, review, and status tools for MCP-compatible agents.
README
Claude Code Plugin for Any Agent
Use an already-installed and authenticated Claude Code CLI from Codex or any other MCP-compatible harness.
This is an MCP server, not a host-specific plugin: it communicates over standard input/output and starts the local claude executable. That makes its tool interface portable across MCP clients while preserving your existing CLI authentication and default model configuration.
What it provides
| Tool | Purpose |
|---|---|
claude_run |
Run a task synchronously or as a managed background job. |
claude_review |
Run a read-only review in a working directory. |
claude_status |
Check local CLI availability and background-job status. |
claude_result |
Retrieve a completed background job's stored output. |
claude_cancel |
Cancel a running background job. |
Every task and review accepts an optional model. If it is omitted, the server does not send --model, so the local CLI uses its normally configured default.
Prerequisites
- Node.js 18 or newer
- Claude Code already installed and authenticated
- Codex CLI, if you are using the Codex installation path below
Check the local prerequisites:
node --version
claude --version
codex --version
Install in Codex
Clone the repository, or use your existing clone:
git clone https://github.com/sarthchawla/claude-code-as-plugin.git
cd claude-code-as-plugin
Register the server. Run this from the cloned repository so the command records an absolute path:
MCP_SERVER_PATH="$(pwd)/bin/claude-code-mcp.mjs"
codex mcp add claude-code \
--env CLAUDE_CODE_MCP_REVIEW_AFTER_RUN=false \
-- node "$MCP_SERVER_PATH"
Verify registration:
codex mcp list
Restart Codex or start a new task. The five tools listed above should then be available.
Enable the post-task review hook
The hook is disabled by default. To enable it, replace the registration:
codex mcp remove claude-code
MCP_SERVER_PATH="$(pwd)/bin/claude-code-mcp.mjs"
codex mcp add claude-code \
--env CLAUDE_CODE_MCP_REVIEW_AFTER_RUN=true \
-- node "$MCP_SERVER_PATH"
The hook runs only after a successful claude_run. It uses a restricted, read-only tool set and returns its findings under review in the same tool response. It adds latency and usage, and does not attempt automatic fixes or feedback loops.
For a customized hook, add either optional environment variable during codex mcp add:
--env 'CLAUDE_CODE_MCP_REVIEW_MODEL=sonnet'
--env 'CLAUDE_CODE_MCP_REVIEW_PROMPT=Review for correctness, security, and missing tests. Do not modify files.'
Use from Codex
Ask Codex to use a tool directly. Typical calls look like these:
{
"prompt": "Investigate the failing tests and make the smallest safe fix.",
"cwd": "/absolute/path/to/project",
"model": "sonnet"
}
Omit model to use the local CLI default:
{
"prompt": "Review the current changes for regressions.",
"cwd": "/absolute/path/to/project"
}
The default permission mode is manual. claude_run also supports permission_mode, allowed_tools, max_budget_usd, timeout_ms, and review_after_run for a single-call hook override.
Background jobs
Set background to true to receive a jobId immediately:
{
"prompt": "Investigate the CI regression and prepare a minimal fix.",
"cwd": "/absolute/path/to/project",
"background": true
}
Then use:
claude_statuswithjob_idwhile the task runs.claude_resultwithjob_idafter status iscompleted.claude_cancelwithjob_idto stop it.
Cancellation sends SIGTERM to the subprocess group, followed by SIGKILL after five seconds if needed. Job state and output are stored only in memory; restarting the MCP server loses prior job history.
Configure another MCP harness
Use the standard mcpServers entry in examples/mcp-server.json, replacing the placeholder path with the absolute location of bin/claude-code-mcp.mjs:
{
"mcpServers": {
"claude-code": {
"command": "node",
"args": ["/absolute/path/to/claude-code-as-plugin/bin/claude-code-mcp.mjs"],
"env": {
"CLAUDE_CODE_MCP_REVIEW_AFTER_RUN": "false"
}
}
}
}
Troubleshooting
- The server is unavailable: run
claude --versionin the same terminal environment used to start Codex. The server relies onclaudebeing onPATH. - The model is not what you expected: omit
modelto use the CLI default, or supply the desired model alias/value only for that invocation. - A background job disappeared: jobs are process-local. Re-run the task after restarting Codex or the MCP server.
- Update or remove the server: use
codex mcp remove claude-code, then run the registration command again (or leave it removed).
Security
Any MCP client that can call this server can ask the local CLI to operate in its supplied cwd. Register it only with trusted harnesses and repositories. Keep the default manual permission mode unless the calling environment provides a suitable sandbox.
Development
npm test
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。