We gave our SaaS an MCP server (~150 tools) — now Claude runs our project management. Lessons learned.

We're building TRCR — time tracking, tasks, clients and invoicing for freelancers, agencies and small teams. The core idea: you track time once, and everything downstream is automatic. Every team member has two rates — what the client pays and what the worker costs you — so revenue, margin and profitability per project/client are calculated in real time, and invoices are generated straight from tracked time entries. No exporting timesheets to a spreadsheet to figure out if a project made money. A few months ago we made a second bet: instead of just shipping features, we exposed almost the entire product as an MCP server so AI agents can use it directly. Today it covers \~150 tools across 21 domains (tasks, CRM, invoices, time, reports, chat...), with OAuth 2.1 and personal access tokens. And we dogfood it hard: our own task management and CRM run through Claude talking to our API. We even hand work to a coding agent by task ID — it pulls the full spec from the task itself. What we learned: **1. Agents are your most brutal API reviewer.** The moment an AI started using our API daily, every inconsistency surfaced — confusing parameter names, endpoints that return too little, missing bulk operations. Fixing the API for agents made it better for humans too. **2. Tool descriptions are product copy now.** An agent picks tools based on the description string. Vague description = tool never gets called. We rewrote them like microcopy, not docs. **3. Self-contained context beats clever integrations.** When we delegate a task to an agent, everything must live in the task itself — full spec in the description, source files as attachments. Any "it can just look it up elsewhere" assumption breaks. **4. OAuth 2.1 for MCP is painful but worth it.** PATs got us started fast; OAuth is what makes it feel like a real integration instead of a hack. **5. Agents + billing data = the killer combo.** The most magical moment wasn't task management — it was asking "which client was most profitable last month and draft them an invoice for June" and watching it pull time entries, apply the right rates, and produce the invoice. That's when exposing the *whole* domain (rates, time, invoices, reports) instead of just tasks paid off. **6. Dogfooding via agents changes your roadmap.** Half our recent backlog came from watching where the agent got stuck. The unexpected outcome: "works well with AI agents" is turning into our main differentiator, not a side feature. Happy to answer questions about building an MCP server for a SaaS — schema design, auth, what I'd do differently.