Codex ChatGPT Web

Codex ChatGPT Web is an open-source, MIT-licensed desktop launcher published by the GitHub user miuuyy. It lets users pick ChatGPT Web models available on their account, including Pro, from Codex's native model picker. Those models run against ChatGPT Web's separate usage limits rather than Codex quota. The latest release listed is v6.1.8, with installers for Windows, macOS and Linux.

What It Is

The tool is unofficial browser automation that bridges Codex and the ChatGPT website. You sign in to ChatGPT in an embedded browser, install the models, restart Codex, and choose a model ending in (Web). Tasks, images and streaming work through the same Codex interface. The app bundles its own browser and runtime, so no separate Chrome, Node or Bun install is needed.

Modes and Models

There are three modes:

• Browser-only sends messages automatically without local Codex tools.
• Full harness (With Automation) sends messages automatically and gives ChatGPT access to the current task's files, terminal, tools and approvals through MCP.
• Zero Risk has the user paste and send the prepared prompt manually, while still using local tools through a separate MCP connector.

Available models depend on the account. Free/Go accounts get Luna / Think. Paid accounts get GPT-6 Sol and GPT-5.6 Sol with Instant to High effort, plus Extra High and Pro when the account exposes them.

Full Harness Setup

Full mode connects ChatGPT's tool calls back to the current Codex task through the official OpenAI tunnel-client. The README states the tunnel is outbound only and needs no public IP, inbound port or router forwarding. The launcher's MCP page walks through creating the Tunnel and API key and connecting the harness. Users then enable ChatGPT Developer Mode, create a Tunnel connector named Codex Native2, and run Verify runtime. Write actions also depend on the workspace administrator policy.

Requirements and Caveats

• It is not an OpenAI API and is not affiliated with OpenAI. Changes to the ChatGPT UI can break selectors, and drift fails explicitly rather than silently switching model or transport.
• Browser state is a sensitive login artifact, so users should keep the launcher profile private and use a trusted workstation.
• Release packages target macOS 13+ (arm64/x64), Windows x64 and Linux x64. Builds are not platform-signed, so Gatekeeper or SmartScreen may warn.
• Terminal installers verify published SHA-256 checksums. The CLI also manages subagent protocols (Compatibility V1 or Native).