ChatGPT Codex Bridge
Self-hosted MCP server that lets ChatGPT work with your local codebase through explicit tools.
README
ChatGPT Codex Bridge
ChatGPT Codex Bridge is a self-hosted MCP server that lets ChatGPT work with your local codebase through explicit tools. It can read, edit, search, run terminal commands, show change summaries, and optionally delegate larger engineering jobs to the local Codex CLI.
The project is designed for people who want ChatGPT to behave more like a local coding agent while keeping the actual source files, terminal, credentials, and tooling on their own machine.
What It Does
- Exposes selected local folders to ChatGPT through MCP.
- Provides workspace-scoped tools for reading, editing, writing, searching, and inspecting files.
- Runs Bash and Windows PowerShell commands for tests, builds, git, package scripts, and local diagnostics.
- Supports permission modes from read-only to full local development access.
- Adds audit snapshots, change summaries, and session rollback for edits made through the bridge.
- Loads project instructions from
AGENTS.mdandCLAUDE.md. - Discovers local skills from bundled and user-configured skill directories.
- Supports ChatGPT Apps-compatible tool cards and workspace UI metadata.
- Can delegate larger tasks to
codex execwith configurable model, reasoning, sandbox, speed, plan policy, and goal policy. - Includes a local settings page for roots, public URL, permission mode, audit, and Codex delegation defaults.
Detailed feature notes are in docs/features.md.
Install
Requirements:
- Node
>=20.12 <27; Node 22 LTS is recommended. - npm.
- Git.
- Bash, such as Git Bash or WSL on Windows.
- A public HTTPS tunnel or reverse proxy when connecting from ChatGPT.
- Optional: OpenAI Codex CLI if you want to use
delegate_to_codex.
Install from a local checkout:
npm install --include=dev
npm run build
node dist/cli.js init
node dist/cli.js serve
After publishing to npm, the same app can be installed globally:
npm install -g chatgpt-codex-bridge
codex-bridge init
codex-bridge serve
The legacy devspace command is also kept as a compatibility alias.
Quick Start
-
Start an HTTPS tunnel to the local server port, usually
7676.cloudflared tunnel --url http://127.0.0.1:7676 -
Initialize the bridge.
codex-bridge initDuring setup, enter:
- the local folders ChatGPT may open;
- the local port, usually
7676; - the public HTTPS origin, without
/mcp.
-
Start the MCP server.
codex-bridge serve -
Add this MCP endpoint to ChatGPT or another MCP host:
https://your-tunnel-host.example.com/mcp -
Approve the connection with the Owner password printed by
init.
The default local settings page is:
http://127.0.0.1:7676/settings
Full deployment instructions are in docs/deployment.md.
Configuration
The bridge stores local config and auth files under:
~/.devspace/config.json
~/.devspace/auth.json
The ~/.devspace path is intentionally retained for backward compatibility.
You can override it with DEVSPACE_CONFIG_DIR.
Common settings:
| Setting | Default | Purpose |
|---|---|---|
DEVSPACE_ALLOWED_ROOTS |
current directory during setup | Filesystem roots ChatGPT may open. |
DEVSPACE_PUBLIC_BASE_URL |
required | Public origin, without /mcp. |
DEVSPACE_PERMISSION_MODE |
safe_shell |
Tool permission policy. |
DEVSPACE_CODEX_DELEGATION |
1 |
Enables delegate_to_codex. |
DEVSPACE_CODEX_DEFAULT_MODEL |
gpt-5.5 |
Default Codex model. |
DEVSPACE_CODEX_DEFAULT_REASONING_EFFORT |
xhigh |
Default reasoning effort. |
DEVSPACE_CODEX_DEFAULT_SPEED |
normal |
Codex service tier, normal or fast. |
See docs/configuration.md for the full reference.
Security Model
This bridge gives a remote MCP client access to local development capabilities. Treat it like granting a trusted coding partner access to your machine.
The safety model is based on:
- a narrow filesystem allowlist;
- an Owner password approval flow;
- OAuth-protected MCP sessions;
- Host header validation;
- permission modes for edits, shells, and Codex delegation;
- audit logs and rollback support for bridge-made file edits.
Do not expose it with broad roots like ~, /, or C:\ unless you understand
the risk. Keep ~/.devspace/auth.json private.
More details are in docs/security.md.
Package
Create a distributable npm tarball:
npm run build
npm pack
The package includes dist, docs, scripts, and README.md.
Local Development
npm install --include=dev
npm run dev
npm run typecheck
npm test
npm run build
npm run start
License
MIT
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。