🚀 What We Learned Building FastAPI-MCP

A few months ago, we released FastAPI-MCP, an open-source library that converts FastAPI apps into MCP servers with zero configuration. We focused on making it easy and intuitive and the response was incredible - developers loved how easy it made MCP integration.

But with usage we started noticing developers had different issues than we originally thought.

The Problem We Didn't See Coming

"Can you help with authentication setup?"

"Can this MCP scale?"

"How do I deploy it to a separate instance?"

"How do I know if my tools are actually working well?"

"How can I tell how my MCP is being used?"

FastAPI-MCP solved the direct conversion problem beautifully. But production? That's where users had to work.

What's The Delta To Production MCP

Here's what we learned:

• Hosting headaches: Where do you even deploy an MCP server? we made it easy to be deployed along side your API but managing two different beasts on the same infra might not be easy.

• Authentication: Bearer tokens are easy, but OAuth takes a lot of set up from devs, even when the package supports it.

• Zero observability: Are agents actually using all the tools? Is your tool description or parameters confusing the LLM? You'll never know

• Scaling issues: What happens when 100 AI assistants hit your server?

API to MCP Conversion Isn't That Easy

Beyond the technical deployment challenges, we discovered deeper issues:

Tool Selection: Developers had APIs with 200+ endpoints but no idea which ones would actually be useful for AI assistants. Should they expose everything? Just the core features? How do you decide? Should we add composite tools to make life easier for the LLMs?

Description Quality: This was huge. Developers would write API documentation for humans, but AI assistants need different context. Tool descriptions that made perfect sense to developers completely confused LLMs.

Connection Complexity: How do customers actually connect to their MCP servers? While FastAPI-MCP creates SSE servers, helping customers integrate with tools like MCP-remote or custom clients was often more complex than building the server itself.

We felt like we're repeating the known pattern "easy to demo, hard to get to production".

The Platform We Wish We Had

Instead of just documenting "how to deploy FastAPI-MCP" we built the infrastructure we wish existed.

Tadata creates hosted, authenticated MCP servers with analytics in under 1 minute. Here's how it works:

# Deploy your FastAPI app
import tadata_sdk
tadata_sdk.deploy(fastapi_app=app, api_key="TADATA_API_KEY")

Or upload any OpenAPI spec - works with Django, Express, Spring Boot, whatever.

📊 The Analytics Game-Changer

Developers can finally see:

• How are agents using my MCP
• Which tool descriptions confuse AI assistants
• Which tools never get used (and should be removed)
• How to optimize based on real usage patterns instead of guessing

Two Different Solutions, Same Vision

Today we maintain both:

• FastAPI-MCP: Open source, DIY hosting, full control
• Tadata: Managed hosting, focus on building tools not infrastructure

They serve different needs in the MCP ecosystem.

FastAPI-MCP remains our love letter to the developer community - free, open, and powerful. Tadata is for developers who want to skip the operational headache and get straight to building great AI tools.

What's Next?

The MCP ecosystem is exploding, and we want to help everyone get the most out of their existing APIs.

If you need production-ready hosting without the hassle, check out Tadata.

The AI revolution is happening. Your software should be part of it 🤖