English | 简体中文
Use WeChat to check ChatGPT Work, continue ChatGPT project conversations, and explicitly trigger read-only Codex analysis on your Windows PC.
WeChat → ClawBridge → ChatGPT Work / ChatGPT Web / Codex CLI
ClawBridge is a small, local-first bridge for developers who want a lightweight mobile entry point to AI work already running on their PC. It is deliberately not a general-purpose agent framework.
/workobserves local ChatGPT Work status and results./chatdiscusses the current project with ChatGPT and lightweight local context, but does not execute code changes./codexexplicitly analyzes a configured repository with Codex CLI in a read-only sandbox.- Detailed work, source code, logs, and long outputs stay on the PC; ClawBridge does not require exposing the development machine directly to the public Internet.
If you only want one part of the workflow, the other capabilities remain optional.
Status: Alpha / Early Stage
The complete first-time setup flow is currently verified on Windows only. ClawBridge is still experimental and is not production-ready infrastructure. Tencent iLink / WeChat ClawBot is an external dependency whose account, session, binding, rate-limit, and delivery behavior may change independently of ClawBridge.
ClawBridge is an independent open-source project and is not affiliated with, authorized by, or endorsed by Tencent, WeChat, or OpenAI.
Ask about the local project directly from WeChat:
For example:
/chat What problem does this project solve?
ClawBridge keeps the conversation tied to the current project context instead of treating every WeChat message as a completely unrelated request.
WeChat ClawBot
↕
ClawBridge
├─ /work → local ChatGPT Work status
├─ /chat → ChatGPT Web + project context
└─ /codex → local Codex CLI (read-only analysis)
↓
local projects
The command boundary is deliberate:
/workobserves local Work facts, status, and results./chatdiscusses the project with ChatGPT and may prepare a Codex handoff./codexexplicitly analyzes a configured local project with Codex in a read-only sandbox.
A /chat message that asks to modify code, run tests, commit, or deploy does not automatically perform those actions.
ClawBridge is intentionally not a general-purpose agent framework. It does not provide model routing, RAG, long-term memory, or workflow orchestration, and it does not depend on QClaw or OpenClaw.
| Capability | Purpose | Requirement |
|---|---|---|
/work |
Observe local ChatGPT Work facts, status, and results | Accessible local Work state |
/chat |
Discuss and analyze the current project with ChatGPT | Chrome with ChatGPT Web signed in |
/codex |
Explicitly start local read-only Codex analysis | Configured project + working Codex CLI |
| Git Context | Add on-demand, read-only Git facts to /chat |
Git repository + Git CLI |
| WeChat transport | Send commands and receive results from WeChat | WeChat + Tencent iLink / WeChat ClawBot |
The complete first-time setup is currently verified on Windows. Linux and macOS are not yet accepted platforms.
Prepare only the dependencies required by the capabilities you plan to use:
- Python 3.11 or later;
- Google Chrome, for the dedicated ChatGPT Web profile;
- a ChatGPT Web account, only for
/chat— ClawBridge does not use the OpenAI API; - WeChat and Tencent iLink / WeChat ClawBot, for WeChat transport;
- Codex CLI, with its own authentication/configuration complete, only for
/codex; - Git CLI and a Git repository, only for Git Context;
- one or more local project directories;
- locally accessible ChatGPT Work state, only for Work Observer.
Cloning the repository and running python -m clawbridge run does not automatically enable every capability. For example, /work does not require ChatGPT Browser, while /chat does.
git clone https://github.com/samzhou1972/clawbridge-wechat.git
cd clawbridge-wechat
python -m pip install -e .Confirm the CLI is available:
python -m clawbridge --helpRun the non-destructive first-run check at any time:
python -m clawbridge doctorThe doctor reports local setup gaps as SETUP rather than treating optional capabilities as failures. Use python -m clawbridge chat-browser doctor when you want to verify the ChatGPT browser session itself.
Runtime configuration is stored at:
%LOCALAPPDATA%\ClawBridge\config.toml
Copy config.example.toml to that location, then explicitly configure every project ClawBridge may use:
[projects.clawbridge]
root = "C:\\path\\to\\clawbridge"Here:
clawbridgeis the logical project name;rootis the real local directory.
[projects.<name>].root is the sole project-path authority for /chat context, Git Context, Work matching, and /codex.
From WeChat:
/chat use clawbridge
selects that configured project.
ClawBridge does not scan the disk to guess your projects. Do not commit personal runtime configuration, credentials, or browser profiles to Git.
ClawBridge uses Tencent iLink / WeChat ClawBot as its transport layer rather than a normal WeChat Web API.
Start the binding flow:
python -m clawbridge loginThe terminal requests and displays a binding QR code. Then:
- scan it with WeChat on your phone;
- if WeChat displays a numeric verification code, enter it in the terminal;
- finish the ClawBot binding flow;
- keep the resulting credentials in local runtime state only.
Run the minimal transport smoke test:
python -m clawbridge echoThen send this to WeChat ClawBot:
hello
Expected reply:
world
This verifies only:
WeChat ↔ Tencent iLink ↔ ClawBridge
Complete this step before diagnosing ChatGPT, Work, or Codex.
API acceptance does not prove that the WeChat client actually received the message. See External service limitations.
ClawBridge uses a dedicated Chrome profile:
%LOCALAPPDATA%\ClawBridge\browser\chrome-profile
Run the one-time login setup:
python -m clawbridge chat-browser setupThis launches normal system Chrome with the dedicated profile.
In the opened browser:
- manually sign in to
chatgpt.com; - confirm ChatGPT works;
- close that Chrome window.
The authenticated session remains in the dedicated profile.
ClawBridge never asks for, collects, or stores your ChatGPT password.
Being signed in to your everyday Chrome or Edge does not mean the ClawBridge profile is signed in. ClawBridge does not use your daily Edge profile or a browser extension.
After closing the login browser, run:
python -m clawbridge chat-browser doctorA ready environment reports PASS for the browser and ChatGPT readiness checks.
Then run the end-to-end browser check:
python -m clawbridge chat-browser doctor --sendOn success:
CLAWBRIDGE_BROWSER_OK
For first-time setup, use the doctor command before diagnosing /chat itself.
If you want to use /codex, install and authenticate Codex CLI separately.
ClawBridge explicitly runs configured projects in a read-only sandbox:
codex exec --sandbox read-only
ClawBridge does not provide a Codex/OpenAI account and does not silently route ordinary /chat messages into Codex execution.
python -m clawbridge runThis starts the long-running bridge process.
If the PowerShell window stops, ClawBridge stops too. If WeChat no longer receives replies, first confirm that the bridge process is still running.
For a first installation using all current capabilities:
- clone and install ClawBridge;
- create
%LOCALAPPDATA%\ClawBridge\config.tomland configure at least one project; - run
python -m clawbridge login; - run
python -m clawbridge echo, sendhelloin WeChat, and expectworld; - run
python -m clawbridge chat-browser setupand sign in to ChatGPT; - close that Chrome window;
- run
python -m clawbridge chat-browser doctor; - run
python -m clawbridge chat-browser doctor --send; - confirm Codex CLI if you plan to use
/codex; - run
python -m clawbridge run; - try
/work status,/chat status, and/chat use clawbridgefrom WeChat.
| Command | Meaning |
|---|---|
/work status |
Show the latest Work status |
/work last |
Show the latest Work activity |
/work result |
Show the latest Work result |
/work watch |
Enable one brief notification when a whole task first reaches COMPLETE |
/work unwatch |
Disable Work completion notifications |
/codex <project> <task> |
Explicitly start read-only Codex analysis for a configured project |
/chat use <project> |
Bind the configured project to the default ChatGPT session |
/chat <message> |
Discuss or analyze with ChatGPT; it does not execute code changes |
/chat status |
Show local chat-session status |
/chat reset |
Reset ChatGPT thread/history while keeping the project binding |
A /chat conversation can continue across messages and keep the selected project context:
The intended workflow is simple:
discuss with /chat, observe with /work, and explicitly use /codex only when local analysis is actually needed.
WeChat is treated as a control, short-summary, and completion-reminder channel, not a long-log channel.
Current policy:
- a business response normally sends at most one WeChat message;
- the outbound hard cap is 1000 Unicode characters;
- automatic multipart / numbered-message chunking is disabled;
/chatasks ChatGPT for a short summary by default;- a Codex handoff of 1000 characters or fewer is provided in full;
- a Codex handoff over 1000 characters is neither truncated nor split; ClawBridge sends a notice to continue on the PC;
- full reports, logs, tracebacks, test output, and detailed development work stay on the PC;
- a proactive Work notification is sent only when a whole task first reaches
COMPLETE, and remains intentionally brief.
Local runtime state lives under:
%LOCALAPPDATA%\ClawBridge
It may contain:
config.toml;chat\session.json;- the dedicated Chrome profile;
- WeChat / iLink account metadata;
- route/runtime state.
ChatGPT login data stays in the dedicated Chrome profile. The iLink token is stored through the local credential store / OS keyring.
Never commit tokens, credentials, profiles, session state, or personal runtime configuration to Git.
Tencent iLink / WeChat ClawBot is an external transport service. Tencent controls availability, rate limits, message volume/frequency, session and context validity, binding state, and actual client delivery behavior. These may change independently of ClawBridge.
In particular:
- Tencent does not publish a stable fixed rate-limit threshold for this use case, so do not rely on a permanent “N messages per minute” rule;
- short bursts or frequent outbound messages may be limited, which is why ClawBridge uses short messages and low-frequency proactive notifications;
sendmessage ret=0or similar API acceptance only means that the server accepted the request — it does not guarantee delivery to the WeChat client;- ClawBridge records such calls as
ACCEPTED_UNCONFIRMEDand does not densely auto-retry delivery that has not been confirmed; - supervised real WeChat E2E delivery has been verified, but account-specific binding or client-delivery anomalies may still occur independently of ClawBridge.
WeChat ClawBot
↕
ClawBridge Core
├─ M1 Work Observer
├─ M2 Codex Executor
└─ M3 Conversation Core
└─ ChatGPT Browser Driver
Runtime browser automation uses Playwright with the same dedicated profile used during manual Chrome login.
It waits for a new assistant response, stable non-empty text, and the corresponding completion signals.
The default session stores its canonical ChatGPT /c/... URL and up to 24 recent audit messages at:
%LOCALAPPDATA%\ClawBridge\chat\session.json
Later runs reopen that exact thread.
ClawBridge does not search the ChatGPT sidebar, inject artificial history, or silently replace an unavailable thread with another one.
Project context is demand-driven and read-only.
It may summarize:
- Work state;
- Git branch;
- working tree clean/dirty state;
- changed-file count;
- last local commit.
It does not perform:
git fetch/pull;- checkout;
- commit;
- source-code search;
- project execution.
Only one /chat request drives the default browser session at a time.
Control commands do not start the browser.
The complete assistant response remains local; WeChat receives only the bounded result defined by the outbound policy.
Current acceptance state:
- M0 — WeChat / iLink Transport: automated acceptance passed;
- M1 — Work Observer: automated acceptance passed;
- M2 — Codex Executor: automated acceptance passed;
- M3-A / B / C / D: automated acceptance passed;
- M0 and M3-D: supervised real WeChat E2E verification completed.
This is still an early-stage project. In particular, the WeChat transport depends on Tencent iLink / ClawBot, so the status above should not be interpreted as a production SLA.
ClawBridge solves a very specific problem:
Once development work increasingly depends on ChatGPT, Codex, and local AI tools, how can you leave the computer and still know what is happening, continue the discussion, and explicitly trigger local analysis when needed?
The answer here is not another full remote IDE, and it is not exposing the development machine directly to the public Internet.
WeChat is already on the phone.
ClawBridge therefore uses it as a lightweight entry point:
use /chat to discuss, /work to observe, and /codex only when local analysis is explicitly required.
The project intentionally stays small and opinionated. If that workflow matches your needs, use it as-is. If not, fork it and adapt the bridge to your own local AI setup.
Issues, bug reports, documentation improvements, and pull requests are welcome.
Before contributing, please read:
ClawBridge is released under the MIT License.
If ClawBridge is useful to you, a ⭐ Star helps other developers with the same problem discover the project.