codex-chatgpt-web

Pick ChatGPT Web — Instant, Medium, High, Extra High, or Pro in Codex's native model picker. The bridge sends the complete Codex task context to a fresh ChatGPT Temporary Chat, attaches images, and streams visible reasoning, tool activity, and Markdown back into the same Codex task.

Codex task ──Responses + SSE──▶ codex-chatgpt-web ──controlled Chrome──▶ ChatGPT
     ▲                                │                                      │
     └──────── native UI, context, images, tracing, and tool lifecycle ──────┘

Highlights

• Native Codex harness. This is the same model-picker, task history, context lifecycle, approvals, sandbox, streaming, tracing, and tool UI you already use in Codex—not a second chat client. Like OpenCodex, it changes the model backend while preserving the native workflow.
• Local-first task sessions. Codex remains the source of truth for task history on your computer. Every browser turn starts in a fresh ChatGPT Temporary Chat and receives the complete accumulated Codex context, so browser chats are not reused across tasks or added to normal ChatGPT history.
• The full Codex harness over MCP. In full mode, Instant through Extra High can use the active Codex task's filesystem, shell, images, approvals, and configured tools/apps through MCP. Calls and real results stay inside the same browser response—nothing is simulated as text.
• Pro stays useful. Pro is the one exception: ChatGPT's current Pro mode does not expose the custom MCP connector this bridge needs. Its native capabilities, including web search and research, remain available. Gather local workspace context with Instant through Extra High, switch to Pro, and Pro receives the complete accumulated Codex task for deeper analysis.
• Fail-closed and manually tested. Model selection, long inline context, images, streaming, visible trace, compaction, native tool rounds, cancellation, and Pro were exercised end-to-end on macOS. UI drift and missing capabilities produce explicit errors rather than silent fallbacks.

Temporary Chat is a ChatGPT privacy mode, not anonymity or local-only inference: prompts are still processed by OpenAI and are subject to the account's settings and OpenAI's Temporary Chat policy. This project is unofficial; users remain responsible for complying with applicable OpenAI terms and workspace policies.

Quick start

Browser-only mode needs macOS, Google Chrome, and a ChatGPT account. It does not need an API key, tunnel, system Node/Bun installation, OpenCodex, or a Playwright browser download.

curl -fsSL https://github.com/miuuyy/codex-chatgpt-web/releases/latest/download/install.sh \
  | sh -s -- --browser-only --acknowledge-unofficial

Sign in through the one Chrome window opened by setup, restart Codex once, and select a ChatGPT Web — … model. Pro appears only when it is available on the authenticated account. Normal use starts automatically after macOS login and does not require another terminal command.

Modes

Every picker entry has one fixed ChatGPT mode. Codex still displays its built-in Effort and Speed rows, but changing them cannot silently change the selected browser model. Pro receives the full context already collected by Codex, but ChatGPT Pro cannot initiate local MCP/tool calls.

The proxy keeps Codex's built-in openai provider and live model catalog. It forwards the official catalog unchanged and appends only its ChatGPT Web entries, so native models, task history, approvals, sandboxing, and tool results remain owned by Codex.

Full harness

Full mode connects ChatGPT's tool calls back to the current Codex task through the official OpenAI tunnel-client. The tunnel is outbound: it does not expose a public IP, open an inbound port, or require router forwarding.

1. Create a tunnel in Platform tunnel settings.

2. Create a runtime key with Tunnels Read + Use in Platform API key settings.

3. Install and import the key:

   curl -fsSL https://github.com/miuuyy/codex-chatgpt-web/releases/latest/download/install.sh | sh
   ~/.local/bin/codex-chatgpt-web tunnel key-import
   

4. Run setup with your tunnel id:

   ~/.local/bin/codex-chatgpt-web setup --full \
     --tunnel-id tunnel_0123456789abcdef0123456789abcdef \
     --acknowledge-unofficial
   

5. While doctor reports ready, attach that tunnel to a ChatGPT connector named Codex Native in ChatGPT connector settings, scan its tools, set the intended action permissions, and restart Codex once.

Write/modify actions require a ChatGPT workspace and admin policy that permit them. OpenAI currently documents those actions for Business and Enterprise/Edu workspaces; personal Pro is limited to read/fetch MCP permissions. See developer mode and MCP apps. Unexpected approval prompts fail closed unless --auto-approve-tool-calls is explicitly enabled; that option clicks Allow once, never a permanent grant.

Operations

codex-chatgpt-web doctor
codex-chatgpt-web service status
codex-chatgpt-web tunnel status        # full mode
codex-chatgpt-web browser check
codex-chatgpt-web login                # refresh an expired ChatGPT session
codex-chatgpt-web uninstall --yes

Setup stores private state under ~/.codex-chatgpt-web, installs versioned launchd services, and journals the previous Codex route so uninstall can restore it. It refuses to replace a different route unless --replace-codex-route is explicit, and refuses to stop or update while a task is still active.

If you stop a Codex task between native tool rounds, no Responses request remains on which Codex can signal cancellation. Abort the retained browser turn without stopping the daemon, then retry the update:

codex-chatgpt-web service cancel-turns

Limitations and security

• This is unofficial browser automation, not an OpenAI API. ChatGPT UI changes can break selectors; drift fails explicitly instead of silently switching model or transport.
• Browser state is a sensitive login artifact. Never share or commit ~/.codex-chatgpt-web/browser.
• The Responses listener is loopback-only, but another process running as the same local user can reach it. Use a trusted single-user workstation.
• Browser turns are serialized to protect one profile and prevent transcript reuse across tasks.
• Managed background installation currently supports macOS only.
• Codex Desktop hardcodes Pro's wire effort as Ultra and always shows a Standard speed row. Those controls do not alter the fixed ChatGPT Web model. Renaming them would require patching the signed Codex app.
• macOS may report that Bun was prevented from modifying apps when Playwright launches installed Chrome. The bridge does not modify Chrome; leaving that App Management permission denied is expected.

Read the complete architecture and security model before enabling full mode. Report vulnerabilities through SECURITY.md.

Development

bun install --frozen-lockfile
bun run verify

verify runs dependency auditing, strict TypeScript checks, harness/MCP/config tests, a relocatable runtime smoke test, and a real headless launch of system Chrome.

• Architecture
• Security model
• Contributing

Credits and disclaimer

Portions of the Responses translation, Codex catalog integration, and browser harness were adapted from OpenCodex under the MIT license. See third-party notices.

This project is experimental, independent software. It is not affiliated with or endorsed by OpenAI, and it must not be used to evade usage limits or access controls. Review OpenAI's current Terms of Use and Services Agreement before public distribution.