Sezzlee SQL Server

Sezzlee MCP

Swagger for agents. Sezzlee turns what you already have — an HTTP backend, a folder of spreadsheets, PDF and XML documents — into tools an AI agent can find and call over the Model Context Protocol.

It comes in two shapes:

• The HTTP catalog embeds in a NestJS backend. It exposes your endpoints as a search-first tool catalog and replays every agent call through your backend's own request pipeline, so your authentication, authorization and validation apply unchanged. For a backend on any other stack, an OpenAPI gateway builds the same catalog from its document.
• Source servers are standalone, read-only MCP servers for local files: Excel, XML and PDF, plus a server that hands bounded language work to a local model.

Status: early. Every package below is on npm; APIs may still change. The spec, the fixture corpus and the NestJS SDK are tested in CI.

Why

A naive MCP adapter turns every endpoint into a tool and forwards calls over the network. That floods the agent's context, loses the caller's identity and silently papers over what it cannot represent. Sezzlee instead:

• shows the agent three meta-tools (search_tools, load_tool, invoke_tool) rather than the whole catalog, so the size of your API does not grow the agent's context;
• decides what a caller can see from the backend's own authorization, and still enforces every call in the backend at invoke time;
• treats every silent resolution — a name collision, an unknown argument, a truncated response — as an error with a code the agent can act on.

The longer argument is on the docs site: why sezzlee is not an OpenAPI adapter.

Quick start

Pick what you need. The servers need Node.js 22+.

Each package README has its own quick start, configuration and limits. To build from source, see CONTRIBUTING.md.

Packages

HTTP catalog

Source servers and their cores

Every package is published from CI with npm provenance; the tarballs are checked and smoke-installed before publication.

Also in this repository

• apps/docs — the documentation site (TanStack Start). Run it with pnpm --filter @sezzlee/docs dev and open http://localhost:5180.

How the HTTP catalog works

1. At startup the SDK reads your framework's route and authorization metadata and builds a catalog of tools, named by fixed rules; a name collision fails startup.
2. An agent calls search_tools, gets compact cards for the tools this caller may use, and load_tool for the full schema of the one it picks, with a version fingerprint of it.
3. invoke_tool composes an HTTP request from the arguments and replays it through your pipeline as the caller. The response is mapped to a fixed error vocabulary and a size budget before it reaches the agent. A call that passes version back is refused with tool_changed if the server answering it holds a different definition, as can happen mid-deploy. A behaviour change behind an unchanged schema is caught the same way once you declare it as a revision, on the operation, its controller or the whole catalog.

The rules for each step are in the spec, and the conformance corpus pins them so the reference implementation and the NestJS SDK cannot drift apart.

Documentation

• Guides and reference: the docs site in apps/docs (tutorials, how-tos, explanations).
• The spec: packages/http/spec.
• What is planned: ROADMAP.md.
• Contributing: CONTRIBUTING.md.

License

MIT — see LICENSE.