AgentBridge

AgentBridge lets a local coding agent ask a user-authorized web AI through a browser extension. Version 0.1 connects Codex to the signed-in ChatGPT webpage running in Microsoft Edge.

Codex --MCP ask_chatgpt(prompt)--> AgentBridge --WebSocket--> Edge extension
      <-- final answer ------------------------------------ ChatGPT webpage

Current status

The v0.1 end-to-end path has been verified with a real ChatGPT webpage:

Codex → ask_chatgpt("1+1等于多少？") → ChatGPT → "2" → Codex

Current boundaries:

• One local Edge extension connection.
• One dedicated signed-in ChatGPT tab.
• One text-only request at a time.
• User confirmation remains enabled for MCP calls.
• ChatGPT DOM changes can require adapter updates.

Quick start on Windows

Create the project environment:

py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
npm install

If your network needs a local proxy, append its temporary proxy option to pip and npm; do not commit proxy credentials or environment files.

Start the local bridge:

.\Start-AgentBridge.ps1

Load the unpacked Edge extension from:

extension

Then open a signed-in https://chatgpt.com/ tab. See EDGE_MVP.md for the detailed installation and test flow.

Use from Codex

The project includes a project-scoped MCP configuration at .codex/config.toml. To use AgentBridge from any local Codex task, register the same Streamable HTTP endpoint globally:

codex mcp add agentbridge --url http://127.0.0.1:8765/mcp

Start a new Codex task after changing MCP configuration. Ask it to use the tool explicitly, for example:

请调用 ask_chatgpt 工具，独立审查这个方案。不要发送密钥、私有文件全文或个人信息。

Repository layout

agentbridge_server/  MCP server and authenticated WebSocket bridge
extension/           Edge Manifest V3 extension and ChatGPT DOM adapter
.codex/              Project-scoped MCP configuration
scripts/             Smoke tests for MCP and bridge round trips
tests/               Python tests and browser-adapter fixture

relay_server/, local_connector/, and web/ are the earlier local-control prototype retained during the transition. AgentBridge is the active path.

Safety model

• Bind AgentBridge to 127.0.0.1 in v0.1.
• Authenticate the extension with a locally generated token.
• Accept extension WebSockets only from a chrome-extension:// origin.
• Treat all web-AI output as untrusted advisory text.
• Never send API keys, passwords, tokens, private file contents, or personal data through ask_chatgpt.
• Never let web-AI output directly authorize shell commands, file writes, or permission changes.

Tests

.\.venv\Scripts\python.exe -m pytest
.\.venv\Scripts\python.exe -m scripts.smoke_agentbridge_mcp
.\.venv\Scripts\python.exe -m scripts.smoke_agentbridge_roundtrip
node --check .\extension\background.js
node --check .\extension\chatgpt_adapter.js
node --check .\extension\content.js

Roadmap

1. Stabilize the ChatGPT adapter with diagnostics and streaming.
2. Add a second web-AI provider through the same adapter contract.
3. Add a second local-agent adapter.
4. Introduce bounded planning and review workflows across agents.
5. Publish a public release after security review, documentation, and license selection.

Security reporting

Do not report credentials, browser sessions, local logs, or security details in public issues. See SECURITY.md.