@minzicat/pi-chatgpt-web-adapter

Use ChatGPT web's gpt-5-5-pro (and gpt-5-5-thinking, gpt-5-5) inside Pi as first-class, switchable models — riding your own ChatGPT Plus/Pro subscription.

gpt-5-5-pro is not exposed by the OpenAI Responses API or the Codex CLI at plan pricing. The only way to drive it from your subscription is the chatgpt.com web app. This adapter does that through a managed browser, then re-exposes it as a local OpenAI-compatible endpoint that Pi consumes like any other provider.

Personal-use tool. It automates your own ChatGPT account on your own machine. Single-account, single-machine. See docs/TOS-NOTES.md.

How it works

Pi session  ──/models chatgpt-web/gpt-5-5-pro──▶  pi openai-completions streamer
                                                       │  http://127.0.0.1:1456/v1
                                                       ▼
                                        pi-chatgpt-web daemon (this package)
                                          ├─ managed Chromium (login + sentinel)
                                          ├─ chatgpt.com/backend-api/f/conversation
                                          └─ OpenAI ⇄ web-protocol translation

• Auth is a real browser login (Google / Microsoft / email — your choice). No HAR import, no pasted tokens. The JWT auto-refreshes from the long-lived session cookie.
• Sentinel/Turnstile tokens are minted by the live page, so the adapter survives ChatGPT frontend rotations far better than pure-HTTP reimplementations.
• Pro turns stream on a separate channel; the adapter polls the conversation to completion so you always get the full answer.

Install

Requires Node ≥ 20.12 and a Chromium-family browser (Chrome / Chromium / Brave / Edge).

npm install -g @minzicat/pi-chatgpt-web-adapter

Quick start

# 1. Log in (opens a browser window once)
pi-chatgpt-web auth login

# 2. Verify
pi-chatgpt-web doctor

# 3. In any Pi session
/models chatgpt-web/gpt-5-5-pro

When installed as a Pi package, the extension registers the chatgpt-web provider automatically and starts the daemon on demand — no manual serve needed.

CLI

Command	Purpose
pi-chatgpt-web auth login	Browser login; captures + persists the session
pi-chatgpt-web auth status	Account + token expiry
pi-chatgpt-web auth refresh	Force a token refresh
pi-chatgpt-web serve [--port N]	Run the OpenAI-compatible daemon (default 1456)
pi-chatgpt-web chat "<prompt>" [--model M] [--effort standard|extended]	One-shot
pi-chatgpt-web doctor	Health: auth, browser, daemon
pi-chatgpt-web install-agent	Install a macOS LaunchAgent for the daemon

OpenAI-compatible endpoints

The daemon serves POST /v1/chat/completions, POST /v1/responses, GET /v1/models, GET /health, GET /doctor. Point any OpenAI client at it:

OPENAI_BASE_URL=http://127.0.0.1:1456/v1 OPENAI_API_KEY=x \
  curl -s $OPENAI_BASE_URL/chat/completions \
  -d '{"model":"gpt-5-5-pro","messages":[{"role":"user","content":"hi"}]}'

Limitations (v0.1)

• Text only — no tool/function calling, images, or attachments yet.
• Stateless — each turn replays the conversation as a fresh tree.
• Pro latency — Pro turns take minutes; the stream sends keep-alives.
• Quota — each Pro turn consumes one Pro use from your plan.

Configuration (env)

Var	Default	Meaning
PI_CHATGPT_WEB_PORT	1456	Daemon port
PI_CHATGPT_WEB_CHROME	auto	Path to Chromium executable
PI_CHATGPT_WEB_HEADFUL	unset	Run the browser visibly (debugging)
PI_CHATGPT_WEB_PROFILE_DIR	~/.cache/pi-chatgpt-web/profile	Browser profile
PI_CHATGPT_WEB_AUTH_FILE	~/.pi/agent/chatgpt-web-auth.json	JWT hot cache (mode 600)
PI_CHATGPT_WEB_LOG_LEVEL	info	debug|info|warn|error

Troubleshooting

• doctor says not logged in → pi-chatgpt-web auth login.
• 401 on chat → session cookie expired; re-run auth login.
• no Chromium found → install Chrome or set PI_CHATGPT_WEB_CHROME.
• Empty answers / sentinel errors after a ChatGPT update → the frontend rotated; restart the daemon. If it persists, the in-page mint may need an update (docs/ARCHITECTURE.md).

License

MIT