chatgpt-mcp

# chatgpt-mcp MCP server that gives ChatGPT's web UI the same `ask(prompt) → response` interface as CLI-based AI tools (Codex, Gemini CLI, etc.). Built with Playwright for browser automation. ## Why this exists This is part of a multi-model orchestration setup where Claude Code acts as the top-level AI orchestrator, dispatching tasks to the best model for the job: - **Codex** (OpenAI) — available via CLI (`codex mcp-server`) - **Gemini** — available via CLI (`gemini-mcp-tool`) - **Claude** — the orchestrator itself - **GPT-5.2 Pro** — only available through ChatGPT's web UI (not API on subscription plans) GPT-5.2 Pro is the most powerful model available for many tasks but lacks a CLI/API interface. This MCP server bridges that gap by automating the ChatGPT web UI via Playwright, giving it the same ergonomic blocking `ask → response` pattern as the other models. ### Use cases - Dispatching complex reasoning tasks to GPT-5.2 Pro from Claude Code - AI round-table discussions where Claude Code facilitates debate between all models - File analysis by uploading documents to ChatGPT - Project-scoped conversations using ChatGPT's Projects feature ## Architecture ``` Claude Code (orchestrator) ├── Codex CLI → mcp-server (stdio) ├── Gemini CLI → mcp-tool (stdio) ├── Claude → native └── ChatGPT → chatgpt-mcp (this project) └── Playwright → Chromium → chatgpt.com ``` The server launches a persistent Chromium browser on first use, maintains login cookies across sessions, and uses multi-strategy DOM scraping to extract responses reliably despite ChatGPT's frequently-changing UI. ## Tools | Tool | Description | |------|-------------| | `chatgpt_ask` | Send prompt, optionally switch model/project, poll until complete, return response | | `chatgpt_reply` | Follow-up in current conversation (no model/project switch) | | `chatgpt_upload` | Upload files + optional prompt, poll for response | | `chatgpt_select_project` | Navigate to a ChatGPT Project by name | | `chatgpt_new_chat` | Start fresh conversation (stays in project if set) | All tools are **blocking** — they return only when the response is ready (or timeout). This matches the ergonomics of Codex and Gemini MCPs. ## Setup ```bash # Install dependencies npm install # Install Playwright's Chromium npx playwright install chromium # Build npm run build ``` ### Claude Code config Add to `~/.claude.json` under `mcpServers`: ```json "chatgpt": { "type": "stdio", "command": "node", "args": ["/path/to/chatgpt-mcp/dist/index.js"] } ``` ### First use On first `chatgpt_ask` call, a Chromium window opens at chatgpt.com. Log in manually once — cookies are persisted to `~/.chatgpt-mcp/user-data/state.json` for future sessions. ## Key design decisions | Decision | Choice | Why | |----------|--------|-----| | Browser engine | Playwright (Chromium) | Full DOM access, self-healing selectors, file upload support | | Tool count | 5 | Orchestrator only needs ask/reply/upload/project/newchat | | Session start | Auto on first call | No separate start_session step needed | | Blocking by default | Yes | Matches Codex/Gemini ergonomics | | Default timeout | 60 minutes | GPT-5.2 Pro can think 20+ minutes | | Completion detection | Multi-indicator + content stability | Most robust: checks stop button, streaming flag, regen/copy buttons, send-enabled state, and 3 consecutive stable content checks | | Response extraction | 5-strategy cascade | Handles ChatGPT UI changes: markdown containers → assistant role → articles → conversation turns → fallback | | Polling | Fibonacci backoff | 2s, 3s, 5s, 8s, 13s, 21s, 30s+ — responsive for quick answers, efficient for long ones | ## Lineage Merges the best of two prototypes: - **gpt-bridge** — battle-tested response extraction, model selection, project management via Playwright - **chatgpt-desktop-mcp** — blocking `ask` tool with Fibonacci backoff (was AppleScript-based, fragile)