T4A AI Chat Platform

Production internal AI chat at chat.time-4-action.com. Next.js + Vercel AI SDK + Auth0, MongoDB token tracking reconciled against Anthropic billing. FastMCP server over SSE.

The shape of the problem

The company needed a chat surface that could reach into its own systems — product data, order status, partner permissions — without shipping that data to a third party. It also needed to track who spent what, because Anthropic bills in tokens and finance bills in euros.

What I built

A Next.js app that speaks the Vercel AI SDK to Claude on one end and to a FastMCP server over SSE on the other. Auth0 handles identity; MongoDB keeps per-user token ledgers that reconcile against the Anthropic invoice at month close.

Lessons

Reconciliation is the feature. The chat is just the UI.
MCP over SSE is faster to ship than MCP over stdio for web workloads, but the reconnection semantics need care.
Token accounting should be done server-side, at tool-call boundaries, not guessed from prompt length.