ChatGPT Web for Codex is a local bridge that adds ChatGPT web models to Codex's native model picker. It sends the current compiled Codex task context to a fresh ChatGPT Temporary Chat inside an embedded browser, attaches images, and streams the visible reasoning, tool activity, and Markdown back into the same Codex task, so the task interface, context lifecycle, tracing, and tool presentation stay in place.
Which entries appear depends on the signed-in account: Free and Go accounts get a single Luna model, and accounts that expose the reasoning selector keep Instant, Medium, High, Extra High, and Pro as their subscription allows. Sequential messages reuse one task-bound Temporary Chat. At the context boundary the retained agent writes a checkpoint before Codex starts a clean chat, and Codex's own history supplies the fallback when that chat was closed.
| Mode | Local Codex tools | Extra setup |
|---|---|---|
| Browser-only | No; Codex shows a warning | None |
| Full harness | Yes for every listed effort, including Pro | OpenAI tunnel plus a ChatGPT connector |
| Zero Risk | Yes; the turn-bound Codex harness stays available | Separate tunnel and connector; you paste and send the prompt yourself |
Full mode routes ChatGPT's tool calls back to the current Codex task through the official OpenAI tunnel-client — an outbound tunnel that exposes no public IP and needs no inbound port or router forwarding — reaching the task's filesystem, shell, images, approvals, and configured tools. Setup runs from the launcher's MCP page and needs a new connector with all actions allowed; write actions also depend on the ChatGPT workspace and its administrator policy. Missing models or tools, and changes to the ChatGPT interface, fail with explicit errors rather than switching route or capability silently.
Temporary Chat is a privacy mode, not local inference: prompts are still processed by OpenAI under the account's settings. The project is unofficial browser automation rather than an API client, so interface changes can break selectors. Browser state is a sensitive login artifact and the loopback listener is reachable by processes running as the same local user. Release packages target macOS 13 and later on arm64 and x64, Windows x64, and Linux x64, and are not yet platform-signed, so the installers verify a published SHA-256 manifest.
Features
- Native model entries: ChatGPT web tiers appear in Codex's own model picker with the task UI intact
- Task-bound chats: sequential messages reuse one Temporary Chat per task, with checkpointed compaction at the context boundary
- Full harness over MCP: ChatGPT tool calls reach the active Codex task's filesystem, shell, images, and approvals
- Zero Risk mode: the bridge prepares and copies the prompt while you choose the model and send it manually
- Cross-platform launcher: one desktop app owns sign-in, model installation, MCP guidance, health checks, and diagnostics
- Isolated browser profile: sign-in and identity-provider windows stay in the launcher's own private profile, with no session copied between browsers
- Fail-closed behaviour: missing models, missing tools, or interface drift raise explicit errors instead of falling back silently
- No API key needed: the packaged launcher requires no model API key, installed Chromium, or system runtime
