Run ChatGPT Web inside Codex like it belongs there

What it does

This project is a local bridge between OpenAI’s Codex CLI and the ChatGPT website. It launches an embedded browser, signs into your account, and forwards complete Codex task contexts into fresh ChatGPT Temporary Chats. Responses—including reasoning, images, and tool traces—stream back into Codex’s native UI as if they were generated by a built-in model. The result is ChatGPT Web impersonating a native Codex model without consuming API credits.

The interesting bit

The unusual angle is the MCP harness: for non-Pro tiers, the bridge wires ChatGPT’s tool calls back through an OpenAI tunnel to Codex’s local filesystem, shell, and approval UI, so the web app ends up driving your local machine. Even ChatGPT Pro is roped in as a read-only analyst that receives the full accumulated task context. It is essentially browser automation elevated to a systems-integration layer.

Key highlights

• Uses Playwright to automate ChatGPT Web in ephemeral Temporary Chats that leave no persistent history.
• Exposes ChatGPT’s tiered models (Instant through Pro) directly inside Codex’s native model picker.
• Full-harness mode routes tool calls through OpenAI’s tunnel-client so ChatGPT can read local files and execute shell commands via Codex’s existing approval UI.
• Ships as a cross-platform desktop launcher with a bundled browser—no separate Chrome, Node runtime, or API key required.
• Fail-closed design: UI drift or missing capabilities throw explicit errors rather than silently falling back.

Caveats

• Unofficial browser automation means a ChatGPT UI redesign can break selectors overnight.
• macOS Gatekeeper and Windows SmartScreen may flag the unsigned release binaries.
• ChatGPT Pro cannot initiate local MCP tool calls; it only receives accumulated context for analysis.

Verdict

Worth a look if you already pay for ChatGPT Plus or Pro and want Codex’s agentic workflow without API usage caps. Skip it if you need guaranteed stability—this is browser automation subject to the whims of OpenAI’s frontend team.