AI agents suck at using APIs. We tried. Here's what we built to fix it.

Valentin Lemort — Mon, 14 Jul 2025 09:54:52 +0000

AI agents suck at using APIs. We tried. Here's what we built to fix it.

We’ve been been building API automation software since 2019 — lots of internal glue code, tools for clients, and APIs on top of APIs.

When LLM agents started getting good, we tried wiring them into our stack.

It didn’t go well.

Why?

Agents don't understand relationships between objects
They fail to understand the specificities of each API if it goes against their training data
They are not even that great at generating valid JSON
They don’t retry on failure.

They need structure. Clear inputs and outputs. Contracts they can reason about.

We looked at Model Context Protocol (MCP) — a spec for exposing actions and resources in a machine-friendly way.

It makes sense. But writing an MCP server by hand is painful.

Too much boilerplate. Too many edge cases. Not fun.

So we built `mcpresso`

It’s a small TypeScript framework to define MCP-compatible servers with:

schema validation via Zod
metadata exposure by default. So the agent knows exactly what it is dealing with
JSON-RPC routing
retries, rate limiting, etc.

Minimal setup. Just define your resources and actions.

It outputs for the agent:

The resources
The tools
The metadata to understand:
- The structure of the API
- The relationships between objects.

Here’s what it looks like:

import { createResource } from "mcpresso"
import { z } from "zod"

export const invoice = createResource({
  id: "invoice",
  actions: {
    get: {
      input: z.object({ id: z.string() }),
      output: z.object({
        amount: z.number(),
        status: z.enum(["paid", "unpaid", "canceled"]),
      }),
      handler: async ({ input }) => {
        const invoice = await fetchFromDB(input.id)
        return {
          amount: invoice.amount,
          status: invoice.status,
        }
      },
    },
  },
})

Then expose it:

import { createMcpressoServer } from "mcpresso"

export const server = createMcpressoServer({
  resources: [invoice],
})

That’s it. You can run this locally, or throw it on Vercel/Cloudflare.

📘 Repo

🌐 LP

Try it

npm install mcpresso zod

Build your server. You're done.

What’s next

We’re using mcpresso as the base for:

a generator that converts OpenAPI specs into MCP servers
an agent that reads docs and generates MCP logic
a secure runtime with RBAC + human-in-the-loop

But right now, we just wanted to make this part not suck.

Got feedback?

This stuff is early. The spec might evolve.

We’re building in the open — and learning as we go.

If you’ve run into similar issues building for agents, let us know.

🗣 GitHub Discussions

DEV Community: Valentin Lemort

AI agents suck at using APIs. We tried. Here's what we built to fix it.

AI agents suck at using APIs. We tried. Here's what we built to fix it.

Why?

So we built mcpresso

Try it

What’s next

Got feedback?

So we built `mcpresso`