About this project

This project is an unofficial desktop launcher that exposes ChatGPT Web models inside Codex's native model picker. According to the README, it lets users work with the ChatGPT Web models available on their account, including Pro, while using ChatGPT Web's separate usage limits rather than Work or Codex quota. The stated goal is to keep the same interface, tasks, images and streaming behavior. Key described capabilities: - Model access: Free/Go accounts are said to get Luna/Think; accounts with reasoning controls get Instant through High, plus Extra High and Pro when available. The launcher reportedly detects what an account can use. - Modes: Browser-only (automatic sending, no local tools), Full harness (automatic sending with local Codex tools through MCP), and Zero Risk (manual paste-and-send with a separate MCP connector, without reading or operating the ChatGPT page). - Full harness: Connects ChatGPT tool calls back to the current Codex task through the official OpenAI tunnel-client. The README states the tunnel is outbound and does not expose a public IP, open an inbound port, or require router forwarding. Setup involves creating a tunnel and API key, enabling ChatGPT Developer Mode, and creating a connector named Codex Native2. - Diagnostics: An Activity view for local diagnostics, a Settings doctor for end-to-end health, options to cancel a retained browser turn or remove the Codex integration, and an optional setting to save chats in ChatGPT history. - Subagents: Compatibility V1 for cross-backend subagents, or Native mode preserving Codex feature settings and enabling Web-to-Web V2 delegation, switchable via CLI commands. - Packaging: The app bundles its own browser and runtime, so no separate Chrome, Node or Bun installation is required. Release packages target macOS 13+ (arm64/x64), Windows x64 and Linux x64. Installers verify published SHA-256 checksums. - Development: Running from source requires Bun 1.4.0, with commands for the launcher, dev chat harness, verification, subagent smoke tests and packaging. Important caveats stated by the README: this is unofficial browser automation, not an OpenAI API, and ChatGPT UI changes can break selectors. Browser state is described as a sensitive login artifact, and the loopback listener is reachable by processes running as the same local user. Builds are not yet platform-signed, so Gatekeeper or SmartScreen warnings may appear. The project states it is independent and not affiliated with or endorsed by OpenAI, and that it does not bypass authentication or access controls. Context limits are said to depend on account type and selected effort, with a measured 90,000-token window for Plus Medium/High and up to 270,000 tokens with experimental 3x context.