MCP BRIDGE · CLOUDFLARE TUNNEL · CAPABILITY-URL AUTH

网页版 ChatGPT 隔空指挥你电脑上的 Codex 和 Claude CodeWeb ChatGPT remote-controls local Codex & Claude Code

ChatGPT 负责想，你的电脑负责动手。以前你得人肉中转：ChatGPT 里问完方案、复制、粘到终端。这座桥把中转自动化——派活、执行、审查、回传，全程只在 ChatGPT 里打字。 ChatGPT plans, your machine executes. No more copy-pasting between tabs — dispatch, execution, review and results flow back automatically over one encrypted tunnel.

3 角色协作roles 5 个 MCP 工具MCP tools FastMCP Cloudflare Tunnel MIT

看数据怎么流 ↓See the data flow ↓ GitHub

01 · DATA FLOW

数据怎么流：一个小团队The data flow — a tiny team

ChatGPT = 项目经理（想方案、派活）｜Claude Code = 干活的工程师｜Codex = 审查的同事。蓝色向下游派活，绿色原路回传结果。 ChatGPT = PM (plans, dispatches) · Claude Code = engineer (executes) · Codex = reviewer. Blue flows downstream, green flows back.

02 · ROLES & TOOLS

三个角色，五个工具Three roles, five tools

AI	强项Strength	短板Limitation
ChatGPT（网页版）	规划、调研、拆需求Planning, research, task decomposition	在云端，碰不到你的电脑Cloud-only — can't touch your machine
Claude Code（本地）	真改文件、真跑命令、执行落地Real file edits, real commands	规划不如 ChatGPTWeaker at planning
Codex（本地）	真改文件、真跑命令、代码审查Real edits, commands, code review	规划不如 ChatGPTWeaker at planning

MCP 工具tool	作用Purpose
`run_codex`	把任务交给本地 Codex 跑（执行 / 审查）Dispatch a task to local Codex (execute / review)
`run_claude_code`	把任务交给本地 Claude Code 跑（执行落地）Dispatch a task to local Claude Code
`read_file` / `write_file`	读写 `BRIDGE_WORK_ROOT` 内的文件Read/write files inside BRIDGE_WORK_ROOT
`list_dir`	列目录，供规划端探索结构List a directory for the planning side

03 · SECURITY

安全警告——务必读完再用Security warning — read before use

⚠ DANGEROUS FLAGS INSIDE

两个执行工具运行时带 --dangerously-bypass-approvals-and-sandbox（Codex）与 --dangerously-skip-permissions（Claude Code）。含义很直接： Both executors run with --dangerously-bypass-approvals-and-sandbox (Codex) and --dangerously-skip-permissions (Claude Code). Meaning: 谁能连到端点，谁就能在你的电脑上跑任意命令、改任意文件。whoever reaches the endpoint can run any command and touch any file on your machine.

🔒 绝不裸奔公网Never expose it raw：至少用 BRIDGE_PATH_SECRET（不可猜密钥 URL），有条件再叠 BRIDGE_TOKEN。: at minimum BRIDGE_PATH_SECRET (an unguessable capability URL), ideally plus BRIDGE_TOKEN.
🚫 绝不分享你的端点地址Never share your endpoint——那等于把你电脑的钥匙发出去。分享请让对方自建。— that's handing over your keys. Share the repo, not the URL.
📁 BRIDGE_WORK_ROOT 把可读写范围锁在一个目录里，别放任整个 $HOME。 confines read/write to one directory — never all of $HOME.
🔐 .env 有密钥，已被 .gitignore 排除——永远不要提交它。 holds secrets and is gitignored — never commit it.

这是个人效率工具，不是多人生产服务。请自担风险。A personal productivity tool, not a multi-user service. Use at your own risk.

04 · QUICK START

五步跑通Five steps to live

STEP 1

克隆 + 装依赖（Python 3.10+，mcp[cli] + uvicorn）Clone + install (Python 3.10+, mcp[cli] + uvicorn)

STEP 2

配置 .env：生成纯 hex 的 BRIDGE_PATH_SECRET；CODEX_BIN/CLAUDE_BIN 指向真身二进制（shell 里的 codex/claude 常是函数包装，用 which 找真身）Configure .env: pure-hex BRIDGE_PATH_SECRET; point CODEX_BIN/CLAUDE_BIN at the real binaries (shell wrappers won't work)

STEP 3

起 server：python -m app.bridge_server --http（监听 127.0.0.1:8000）Start the server: python -m app.bridge_server --http (127.0.0.1:8000)

STEP 4

起隧道：cloudflared tunnel --protocol http2 --url http://localhost:8000（http2 是必须的，默认 QUIC 过夜会失活）Start the tunnel with --protocol http2 (QUIC default dies overnight)

STEP 5

ChatGPT → Settings → Connectors → 开发者模式 → 添加 https://xxx.trycloudflare.com/mcp/<密钥>ChatGPT → Settings → Connectors → Developer mode → add the secret URL

05 · PITFALLS

踩坑速查——都是真踩过的Pitfalls — all really hit

现象Symptom	原因 / 解法Cause / fix
`run_codex` 报fails with `exit=-6` + dyld `libllhttp.x.dylib not loaded`	Homebrew 升级拽断了 node 的 llhttp 依赖 →upgrade broke node's llhttp → `brew reinstall node`
ChatGPT 声称「被安全检查拦截」但其实没真调claims it was "blocked" without calling anything	没真发工具调用，自己脑补借口 → 措辞强制点名工具 + 要求原样回贴返回It never sent the call → force name the tool + demand verbatim output echo
隧道过夜失活，报Tunnel dies overnight with `quic: timeout`	quick tunnel 默认 QUIC/UDP → 必须加defaults to QUIC → must add `--protocol http2`
ChatGPT 访问端点gets 404	密钥路径含secret path contains `_`/`-` 被 URL 规范化吞掉 →normalized away → 必须纯 hex（pure hex via `secrets.token_hex`）
调本地 codex/claude 没反应Local codex/claude doesn't respond	shell 里是带审批音效的函数包装 → .env 指向真身二进制shell wrappers with approval hooks → point .env at real binaries
trycloudflare 地址重启就变address changes on restart	quick tunnel 是临时的 → 配命名隧道绑自己域名is ephemeral → use a named tunnel with your domain

06 · FAQ

常见问题 · 四问速览FAQ · The four questions

Q.1解决什么问题？What problem does it solve?＋

三个 AI 各关在各的盒子里：ChatGPT 会想但碰不到你的电脑；本地 AI 能动手但规划弱。以前靠人肉中转（复制粘贴），这座桥把中转自动化——ChatGPT 直接派活，结果自动回传。Each AI is boxed: ChatGPT plans but can't touch your machine; local AIs execute but plan worse. The bridge automates the human relay between them.

Q.2什么场景用？得到什么结果？Scenario and outcome?＋

在 ChatGPT 里说「写个三步计划，再让 Codex 审一遍」：Claude Code 真在你硬盘上生成文件，Codex 真读了文件写下审查意见，结果回到对话里。全程只在 ChatGPT 打字。Say "draft a 3-step plan, have Codex review it": files really appear on disk, Codex really reads and reviews, results land back in the chat. You only ever type in ChatGPT.

Q.3内部是什么结构？How is it structured?＋

四个关键招数：① ChatGPT 开发者模式（MCP 入口）；② 本地 FastMCP 桥（一套代码 stdio/--http 两用）；③ Cloudflare Tunnel（不裸奔公网）；④ 密钥路径认证（capability URL——开发者模式发不了自定义请求头，Bearer Token 用不了，改用 /mcp/<hex密钥>）。Four moves: ChatGPT developer mode (MCP entry); a local FastMCP bridge (stdio/--http dual mode); Cloudflare Tunnel; capability-URL auth (developer mode can't send custom headers, so Bearer is out).

Q.4能复用什么？What can you reuse?＋

① FastMCP 桥模式（stdio/--http 一套代码两用）可直接抄；② capability-URL 认证适用于一切「云端只能发 GET 式 connector」的场景；③ 六条踩坑速查能帮任何走 Cloudflare Tunnel + MCP 路线的人省一晚上。MIT。The FastMCP dual-mode bridge, capability-URL auth for header-less connectors, and six hard-won tunnel/MCP pitfalls. MIT.