Ansh Pathak, a high school senior in Texas, has successfully deployed Stopover, a remote Model Context Protocol (MCP) server that integrates with both Anthropic’s Claude and OpenAI’s ChatGPT. The service optimizes multi-city travel itineraries by calculating the cheapest flight or train orders for 2 to 7 cities within a specified date window. Pathak’s implementation highlights the fragile and often undocumented technical constraints imposed by major LLM providers when interacting with external tools.

Navigating Asynchronous Timeouts and Compute Limits

The most critical engineering challenge was reconciling the disparate patience thresholds of MCP clients. While Anthropic’s documentation permits tool calls to run for 240 seconds, Pathak set a conservative 200-second hold for Claude to ensure response delivery. In stark contrast, OpenAI’s ChatGPT lacks official documentation on timeout limits, with community reports suggesting a hard cap around 60 seconds. Pathak discovered through live testing that ChatGPT abandons calls exceeding 46 seconds, forcing him to restructure the plan_trip tool to return a status of "running" with a plan ID after just 25 seconds, relying on subsequent get_plan calls to poll for completion.

Infrastructure Pitfalls on Cloud Run and Vercel

Hosting the server on Cloud Run introduced a subtle but fatal flaw for asynchronous jobs: CPU availability is tied strictly to open request windows. Pathak noted that fire-and-forget jobs crawl to a halt if the HTTP request closes prematurely. To mitigate this, the server keeps the request open during initial processing and maintains it during polling phases to ensure CPU continuity. Additionally, the MCP endpoint could not be hosted on Vercel due to its 300-second function timeout, necessitating a move to Cloud Run with careful state management. The server remains stateless per request, using a database to persist only the plan ID between polling calls.

Implementing OAuth 2.1 and Dynamic Client Registration

For personalized features like saved trips and price-drop alerts, Pathak implemented OAuth 2.1 with dynamic client registration and PKCE via Supabase. This was essential because ChatGPT does not support custom headers or API keys for connectors. The implementation required strict adherence to RFC 9728 for protected resource metadata, ensuring ChatGPT correctly identified the authorization server. Pathak also emphasized the importance of scope isolation, noting that tokens issued to a connector must never function as general site login credentials to prevent security leakage.

Client-Specific Tool Descriptions and Registry Compliance

The project revealed that tool descriptions are not universal; they must be tailored to each client’s discovery mechanism. Anthropic’s review process demands concise, factual descriptions without promotional language or cross-tool instructions. Conversely, OpenAI’s guidelines require descriptions to start with "Use this when..." to help ChatGPT’s selection logic. Pathak maintains two sets of descriptions for the same tools, switching based on the calling endpoint. Furthermore, publishing to the official MCP Registry required domain verification via Ed25519 key pairs and strict versioning, with the server domain needing to match the service domain for Anthropic directory inclusion.

Key Takeaways

  • ChatGPT enforces an undocumented ~46-second timeout for MCP tool calls, significantly shorter than Claude’s 240-second limit.
  • Cloud Run’s request-bound CPU model requires keeping HTTP connections open during asynchronous job processing to prevent stalling.
  • OpenAI’s ChatGPT connectors require OAuth 2.1 with dynamic client registration, as they do not support custom API keys.
  • Tool descriptions must be client-specific, with Anthropic preferring factual brevity and OpenAI requiring contextual "Use this when" phrasing.
  • MCP Registry publication demands domain ownership verification via Ed25519 keys and strict version bumping for every update.

The Bottom Line

Pathak’s work proves that building for the MCP ecosystem is less about model intelligence and more about navigating the brittle, undocumented infrastructure constraints of proprietary clients. The divergence in timeout and description requirements between Claude and ChatGPT suggests that developers must maintain dual implementations from day one, or risk their integrations failing silently in production.