mcp-creator

Choose one of three archetypes based on the target: web scraping with Axios and Cheerio for public read-only content, Puppeteer browser automation for JavaScript-heavy interactive sites, or private-API reverse engineering of internal XHR endpoints for authenticated portals. The scaffolding script generates a starter project for each archetype.

ChatGPT requires a remote HTTPS SSE endpoint and does not support local stdio transport. Expose standardized search and fetch tools, set readOnlyHint: true on all read operations to avoid approval prompts, wrap Resources and Prompts as callable tools, and tunnel the server via ngrok or Cloudflare Tunnel.

The most common cause is stdout pollution: any console.log or print writing to stdout corrupts the JSON-RPC 2.0 stream in stdio transport. Route all diagnostic logging to stderr, then verify with the MCP Inspector or the verify_mcp_server.py script which detects non-JSON stdout output.

The Meta-Tool pattern registers only four tools (list_plugins, search_tools, get_tool_schema, call_tool) instead of exposing every upstream tool schema. This reduces context consumption from roughly 40,000 tokens to about 1,200 tokens when aggregating 50-100+ tools across multiple servers.

Use stdio for local desktop clients like Claude Desktop, Cursor, and VS Code since it needs no network ports and runs as a 1:1 subprocess. Use Streamable HTTP with SSE for remote clients, ChatGPT, or shared team infrastructure, with HTTPS and Bearer token authentication enforced.

Browser automation has high latency (5-30 seconds), lower reliability due to DOM changes and timing issues, and struggles with MFA or CAPTCHA challenges. It is best suited for prototyping; private-API reverse engineering offers faster, more stable production integration when endpoints can be captured.