MCP Server for Reportia

The problem

Reportia, my own product under CtisEvolution, exposes an HTTP API for companies, accounting transactions, commissions, account mappings, and SIIGO sync. Giving that API to an AI agent the most direct way — a generic tool like http_request(url) — means the model can point the client at any URL, including an internal IP. It's the classic entry point for an agent-induced SSRF.

The approach

The MCP server exposes 66 curated tools, one per real business endpoint, instead of an open surface. Authentication supports three modes — a pre-issued cookie, a bearer token, and email/password login — with precedence cookie > bearer > session, because the cookie already matches the identity resolved upstream. If REPORTIA_TOKEN and REPORTIA_COOKIE are set together, the server deliberately throws a configuration error, so a token set by accident doesn't override an already-valid per-user session. There's a dedicated test file that verifies the order in which security guards are applied — for example, that the company is resolved first and only then is confirmation of a destructive operation validated — not just that they exist, so a configuration error returns a useful message instead of a generic rejection.

Result

Published on npm and in the official MCP Registry. Any automation over Reportia stays bounded to a surface of 66 known, auditable operations, instead of a generic tool an agent could misuse — with no extra effort from whoever uses it.