31 posts tagged with "MCP"

If you've used AI assistants, you've probably hit this frustration: you need to re-explain context every time you start a new conversation. That database schema you carefully described yesterday? Gone. The code snippet you asked about last week? You're pasting it in again.

MCP Resources solve this. They let you give your AI access to files and data that persist across conversations — like giving your assistant a file cabinet instead of asking them to memorize everything.

Following our prior post on wiring up an MCP server for our Django app — see How We Integrated Model Context Protocol (MCP) into Our Django App — we went back and revisited the architecture. "Too many tools" is still a huge problem for LLM productivity, which has continued into GPT5 and the latest Claude models so probably won't be solved toon. Cursor and Claude both work better when they have fewer tools to choose from, and our original setup exposed too many single-purpose GET tools. So we consolidated everything into a single, strongly-typed batch tool.

The result: one get tool, clearer schema, faster concurrent fetches, and less model confusion.

OpenAI just made it clear: if you want your MCP server to work with ChatGPT's deep research capabilities, it needs to implement two specific tools: search and fetch.

This isn't optional. It's the protocol standard that enables ChatGPT to discover and retrieve information from external data sources. And if you're using MCPBundles, you already have it.

Most "MCP integrations" ship as thin wrappers over existing REST endpoints because it's the fastest way to get something working, but that convenience often fights how agents actually reason and act. If you want MCP that agents can use reliably, design it like an action‑oriented RPC interface rather than a collection of CRUD‑style endpoints.

Most APIs weren't built for AI agents. Stripe has 300+ endpoints. GitHub's API returns paginated results that change between requests. Slack rate-limits you after 20 calls per minute. None of them were designed for Claude to use directly.

We wrapped dozens of REST APIs into MCP tools and learned that a thin wrapper doesn't work. You need to rethink the interface completely—actions instead of CRUD, consolidated search instead of endless list endpoints, and server-side resilience for pagination and rate limits.

Here's what actually works.

Users would ask Claude to "set up all my integrations," and Claude would call our provisioning tool. Then nothing. Users waited 45 seconds staring at a spinner while our server created API keys, configured webhooks, and set up OAuth clients. Most users gave up after 15 seconds, thinking it failed.

Then someone asked Claude to "clean up old bundles," and Claude dutifully deleted everything from the last 6 months. Because we let it.

We needed streaming for long-running work, chunking for large operations, and approval gates for anything scary. Here's what works.

Our bundle search tool was taking 12 seconds at P95. Users would ask Claude to "find the Slack integration," watch nothing happen, then ask again. Claude would make the same call twice, wait 24 seconds total, and users would close the tab thinking our service was down.

The tool worked perfectly. It just worked slowly. And in the world of AI assistants, slow might as well be broken.

Here's the thing about schemas: they're basically the contract your model learns from. Get them right, and your tools are easy to find, hard to break, and simple to fix when something goes wrong.

Get them wrong, and you'll spend way too much time debugging why the model keeps calling your tool incorrectly.

I spent two hours staring at Claude Desktop wondering why my MCP server wouldn't connect. The config looked perfect. The server ran fine standalone. But Claude showed no tools. The problem? A single trailing comma in my JSON config that Claude silently ignored.

This is everything I wish the docs had told me upfront.

New to MCP? Start with Introduction to MCP to understand what it is and how it works.

My first MCP server took three hours to get working because I made every possible mistake: no logging, broke stdio with print statements, forgot to restart Claude Desktop, and wondered why nothing worked. Your first one should take 30 minutes.

This is what actually works, with the debugging steps I wish I'd known upfront.