Use ChatGPT Web as a model backend inside the native Codex task harness.
Korean-first AICC fork with reversible routing and project-bound tools.
Codex keeps the native task UI, history, approvals, tracing, compaction, and tool lifecycle. Only the selected Web model's inference is performed by the signed-in ChatGPT Web account.
This fork is based on upstream
miuuyy/codex-chatgpt-web v2.0.0
at commit 8ae34eb. The upstream attribution and MIT license are preserved.
Codex task -> local Responses bridge (17841) -> embedded ChatGPT Web
| non-Web model
+-----------------------------> preserved loopback upstream
Eight non-Pro entries are always available after account verification:
Web / Temporary / Low,Medium,High,Extra HighWeb / Saved / Low,Medium,High,Extra High
The Korean launcher displays these as Web / 임시|저장 / 낮음|중간|높음|매우 높음.
Extra High (xhigh) is not Pro-only. Two Pro entries are added only when the signed-in account
actually exposes Pro. Conversation retention and reasoning level are fixed in the model slug, so
Codex's separate effort row cannot silently change the selected browser mode.
Each Codex turn currently uses a fresh ChatGPT conversation. Temporary turns do not appear in normal ChatGPT history; Saved turns do. Codex remains the durable history source and sends a verified accumulated context packet for each turn. Reusing one ChatGPT conversation URL across multiple Codex turns is intentionally not implemented yet.
On first setup the bridge detects and preserves an existing loopback Codex upstream and model catalog. On this AICC workstation the resulting topology is:
Web models -> 17841 -> ChatGPT Web inference
Non-Web models -> 17841 -> OCX 10100
codex-native profile -> official Codex endpoint
A Web request does not invoke OCX model inference. If OCX stops, the Web branch remains usable.
If the 17841 front door stops, every Desktop request configured through that front door fails.
Native recovery remains a separate profile or a deliberate ocx restore at a safe restart point.
macOS or Linux
curl -fsSL https://github.com/AgenticLab-SH/codex-chatgpt-web/releases/latest/download/install-launcher.sh | shWindows PowerShell
irm https://github.com/AgenticLab-SH/codex-chatgpt-web/releases/latest/download/install-launcher.ps1 | iexThen:
- Sign in with the intended ChatGPT account inside the app-owned browser profile.
- Run the browser smoke test.
- Install the models. Existing Codex routing is journaled and replaced reversibly.
- Fully restart Codex Desktop once to refresh its model catalog.
The launcher can start at login and keep the bridge alive after its window closes. Web models are unavailable when the Mac is off or the launcher is fully quit.
| Mode | ChatGPT Web inference | Local files, shell, and tools | Extra setup |
|---|---|---|---|
| Browser-only | Yes | No | Embedded browser login |
| Full harness | Yes | Within the current Codex project boundary | Secure MCP Tunnel + ChatGPT app |
Full mode exposes dedicated project-bound file, patch, command, and image tools. Additional
Browser, Chrome, Computer Use, MCP, and app tools are available only when the active Codex harness
advertises them and the configured allowlist permits them. Raw gateways that could bypass the
project boundary, such as unrestricted node_repl, are not exposed directly.
The official Secure MCP Tunnel is outbound-only. It does not require a public IP, inbound port, or router forwarding. A regular API key with Tunnel Read+Use permission authenticates the tunnel control plane; it does not create model API calls. ChatGPT subscription and workspace policies still apply.
The AICC Workspace Business app is a separate path for using registered projects from external
ChatGPT Web, including a phone browser while the Mac, local MCP server, and tunnel are running:
External ChatGPT Web -> Secure MCP Tunnel -> AICC Workspace MCP -> registered project
It does not delegate to Codex tasks and does not consume local Codex or OCX model tokens. Full custom MCP behavior in native mobile apps is not an official guarantee; verify the mobile web surface first.
- Up to five browser turns may run concurrently; completed turn tabs are released immediately.
- Native Codex compaction requests are supported.
- Project traversal, symlink escape, arbitrary absolute paths, and unallowlisted tool calls fail closed.
- Approval prompts are not auto-clicked by default. The explicit opt-in handles one-time approval, not permanent grants.
- ChatGPT UI drift produces an explicit selector/capability error instead of switching accounts, models, or transports silently.
- Temporary Chat is still processed by OpenAI. See the Temporary Chat FAQ.
git clone https://github.com/AgenticLab-SH/codex-chatgpt-web.git
cd codex-chatgpt-web
bun run verify
bun run app
bun run app:packageDevelopment requires Bun 1.3.14. See the architecture, security model, and contributing guide.
This project is independent software and is not affiliated with or endorsed by OpenAI. Use it only with your own authorized account and workspace, in accordance with applicable OpenAI terms and organization policies.