codex-chatgpt-bridge

Languages: English | Japanese

A local experimental bridge that lets Codex ask ChatGPT Web for compact second opinions.

Codex stays the executor. ChatGPT stays the advisor. The bridge only moves small structured delegation packets between them.

Codex -> cgpt CLI -> local browser bridge -> ChatGPT Project -> structured response -> Codex

Status

This is an experimental, unofficial alpha tool.

It is not affiliated with, endorsed by, or supported by OpenAI. ChatGPT Web automation can break when the web UI changes. Use it for local experimentation only.

Important Safety Notice

This project automates ChatGPT Web through a logged-in browser session. Before using or publishing derivatives, review the terms that apply to your ChatGPT account and use case. OpenAI's terms include restrictions around automated or programmatic extraction of Output.

This tool is not intended for scraping, bulk extraction, dataset generation, account sharing, or bypassing API access.

Recommended guardrails:

Why

Codex is strong at acting on the local machine: reading files, editing, running commands, and verifying changes. ChatGPT is often useful as a thinking partner for planning, research, critique, and summarization. This bridge explores a middle path: Codex sends a small delegation packet to ChatGPT and only reads back a short structured result.

The goal is to reduce Codex context usage for compact second opinions without giving ChatGPT control over your machine.

Current Capabilities

Current Limitations

• ChatGPT Web automation depends on the current web UI and can break without a code change in this project.
• Playwright checks the configured Project URL before and after each response. Project-name targeting remains less reliable; use a Project URL when possible.
• Invalid ChatGPT response schemas trigger one formatting-only repair attempt. The original answer is saved as .cgpt/responses/<job-id>.original.md; a failed repair is also saved for inspection.
• Common high-confidence credentials in the question or context block delegation before the job is saved or sent. Detection is best effort; review context packets for less recognizable secrets and sensitive data.
• There is no Chrome extension adapter yet; Playwright is the only automated browser adapter.
• Model selection depends on the web menu and requires the exact visible model label. Labels and available options vary by account and UI locale. It verifies the UI selection, not the backend model identity.
• Headless mode may be blocked by ChatGPT verification. Minimized Chrome has been tested on Windows; other platforms and browser channels may behave differently.

Install

npm install
npm run build

Requirements:

• Node.js 22+
• Chrome or Edge
• A ChatGPT account you can log into locally

First Login

Use a dedicated browser profile for the bridge. Login always opens a visible window; complete any sign-in or verification manually:

node .\dist\cli.js login --channel chrome

If Chrome is not installed:

node .\dist\cli.js login --channel msedge

The browser profile is stored outside the repository by default:

~/.codex-chatgpt-bridge/browser-profile

Override it if needed:

$env:CGPT_BROWSER_PROFILE_DIR="C:\path\to\profile"

Recommended: Use a ChatGPT Project

Create a dedicated ChatGPT Project, for example Codex Bridge.

Save the project URL:

node .\dist\cli.js project-set --url "https://chatgpt.com/g/g-p-.../project"

Or target the project by sidebar name:

node .\dist\cli.js project-set --name "Codex Bridge"

Generate project instructions:

node .\dist\cli.js project-instructions

Paste .cgpt/project-instructions.md into the ChatGPT Project instructions. This tells ChatGPT that incoming messages may be delegated by Codex rather than typed by the human user.

Usage

Ask through the Playwright adapter:

node .\dist\cli.js ask --adapter playwright --mode review --question "List the top 3 risks in this bridge design."

Use a one-off project URL (recommended):

node .\dist\cli.js ask --adapter playwright --project-url "https://chatgpt.com/g/g-p-.../project" --mode plan --question "What should be built next?"

Alternatively, target the sidebar name:

node .\dist\cli.js ask --adapter playwright --project-name "Codex Bridge" --mode plan --question "What should be built next?"

Create a manual prompt packet:

node .\dist\cli.js ask --adapter manual --mode research --question "What is the smallest useful architecture?"

Save a manual response:

node .\dist\cli.js save --job <job-id> --from-file .\answer.md

Read a response:

node .\dist\cli.js show --job <job-id>

Select a model

Use the exact label shown in ChatGPT's model menu:

node .\dist\cli.js ask --adapter playwright --model "GPT-5.6 Sol" --question "Summarize the main tradeoffs in this design."

The bridge checks the selected menu item before submitting and again after receiving the response. It fails if the requested model is unavailable or cannot be verified; it never substitutes another model. Auto is not supported.

You can use --model Latest to select ChatGPT's latest model option. It is a dynamic label, not a fixed model version. The result reports both the requested menu option (model: Latest) and the selector's display text (for example, model_display: 6 Pro). The display text must remain the same before and after the request. This records the UI state for that run; it does not establish a permanent mapping or independently identify the backend model.

For a model that exposes the Power control, --power selects a one-based level from lowest to highest. The number of levels and their labels vary by account and UI locale; the requested level must exist in the current control. A five-level English interface uses:

node .\dist\cli.js ask --adapter playwright --model "GPT-5.6 Sol" --power 3 --question "Review this plan."

The bridge checks the Power control's value and accessible description before and after submission. An unavailable level or inconsistent UI produces POWER_UNVERIFIABLE. --power requires an explicit --model; omit Power to preserve the current setting. Labels and availability can change with the account or web interface. Choosing a level changes the dedicated profile's current Power setting.

Without --model, the browser's default selection is used and reported as unverified. Available labels depend on your account and the current ChatGPT interface.

Browser window options

Playwright runs normal Chrome minimized by default, using the dedicated profile. No window flags are needed. Only run one command at a time against that profile.

--minimized true and --headless true cannot be combined. Login always remains visible. A minimized window can briefly appear during startup or when the site requests interaction.

If ChatGPT requires verification, restore the window or run login and complete it manually. The bridge stops on verification pages, HTTP errors and rate limits; it does not automatically switch modes or bypass these checks.

A response timeout or a model-verification failure after submission can mean the request already ran. Check the chat before retrying. Partial responses are not reported as completed.

Doctor

Run local checks without sending a prompt to ChatGPT:

node .\dist\cli.js doctor

Run browser and Project reachability checks:

node .\dist\cli.js doctor --adapter playwright

The Playwright doctor opens ChatGPT with the configured browser profile, verifies that the prompt editor is reachable, and checks the configured Project target when one is set. It does not submit a delegation prompt.

Modes

Response Contract

ChatGPT responses are validated before they are saved:

verdict: proceed | revise | blocked

summary:
- concise bullet

risks:
- material risk only

sources:
- optional URL

next_action: one concrete sentence

If the response does not include a valid verdict and at least one summary item, the CLI fails instead of saving an ambiguous result.

Adapters

MCP Server

Build the project and run the stdio MCP server:

npm run build
node .\dist\mcp.js

The server exposes:

The Playwright adapter also accepts these optional fields in chatgpt_delegate:

These browser options require adapter: "playwright". The manual adapter remains the default.

Debugging

Debug commands can expose account names, chat titles, project names, and page content. They are gated:

node .\dist\cli.js debug-page --unsafe-debug
node .\dist\cli.js debug-submit --unsafe-debug --text "hello"

Use these only in a private local environment.

Show the active browser profile path:

node .\dist\cli.js profile-path

If ChatGPT appears to "forget" login, first check that cgpt, cgpt-mcp, and any manually opened browser window are using the same profile directory. The default is:

~/.codex-chatgpt-bridge/browser-profile

Codex Skill

The included skill lives at:

skills/chatgpt-delegate/SKILL.md

It tells Codex when to delegate, how to keep context small, and how to treat ChatGPT output as non-authoritative advice.

Local State

Development

npm run check
npm test

Tests include a local model-menu fixture and require Chrome, or Edge selected through CGPT_BROWSER_CHANNEL=msedge. They do not sign in to ChatGPT or send prompts.

Roadmap

• Chrome extension adapter for more stable DOM integration.