MCP in production: lessons from six integrations
The Model Context Protocol gives assistants a standard way to call tools and read context. After six production MCP servers on top of enterprise systems, here is what it changes, and what it doesn’t.
What MCP actually standardises
MCP defines how a model discovers tools, calls them and receives results. That is the whole contract. It does not define authentication, authorisation, rate limits, data residency or what happens when the tool returns forty thousand rows. Every one of those became an engineering task on every one of our six servers.
Lesson 1: the tool list is the product
A model with sixty tools behaves worse than a model with twelve. We ended up designing each MCP server like a public API: a small set of coarse-grained, well-named operations with clear descriptions, not a mirror of the underlying system’s endpoints.
Lesson 2: identity has to flow through
The assistant acts on behalf of a person. The MCP server has to act on behalf of that same person, with that person’s permissions, and the backing system has to see it that way in its audit log. Getting this right took longer than building the tools.
Lesson 3: results need shaping, not just returning
Raw records from an ERP or a fleet platform are not model-friendly. We shape results into short, typed summaries with stable identifiers, and give the model a way to drill down. Token cost fell, answer quality rose.
Lesson 4: MCP is an integration pattern, not a replacement for one
Behind five of our six servers sits the same integration layer we would have built anyway: APIs, events, a canonical model. MCP is the last mile to the assistant. If the middle is missing, MCP just exposes the mess faster.
Treat an MCP server like a product API with a very impatient consumer.
Where it shines
Internal assistants for operations teams, where the questions are repetitive and the systems are known. Read-heavy workflows. Draft-and-confirm actions. In all of those, MCP removed months of custom glue and gave us one place to govern what the assistant can touch.
By Bassem Fawaz · #mcp #ai-agents #integration