MCP for Operators: Data as Tools
Why the Model Context Protocol turns every product and dataset into a callable tool.
The Model Context Protocol is usually explained to developers. The operator version is shorter and more useful: MCP is how a dataset stops being something you export and becomes something a model can call.
The problem it removes
Before, connecting a model to your systems meant one of two bad options. You either pasted data into a prompt — losing freshness, blowing context, and capping how much the model could see — or you built a bespoke integration per model per tool, and maintained it forever.
MCP makes the tool the stable thing. You describe what a capability does and what it needs, once. Any client that speaks the protocol can use it.
What that means in practice
It means your catalogue is a tool. Your CRM is a tool. Your analytics warehouse is a tool. The question stops being "how do I get this data into the prompt" and becomes "what is this system allowed to do".
That second question is the interesting one, and it is an operator question, not an engineering one.
- Read and write are different decisions. In my own server, read tools are always on; writes are gated behind an explicit flag, and destructive operations require confirmation. That is a policy, expressed in code.
- Credentials scope the surface. Each tool group activates only when its credentials are present, so an environment without a key simply does not expose that capability rather than failing at call time.
- The tool description is the contract. If it is vague, the model will use it wrongly, and that is your fault rather than the model's.
The part people underrate
Once tools are callable, the interesting systems are not chat interfaces. They are unattended: a scheduled job that reasons over your own data through your own tools and acts. The chat window is the demo. The loop is the product.
Where to start
Pick the dataset that people ask you questions about most often, and expose reads of it. Not writes. Live with it for a fortnight and watch what gets asked — that tells you what the second tool should be, and it is almost never what you would have guessed.
Keep reading
A Run Produces Data. A Loop Produces Decisions.
The signature thesis: the data you scrape should plan the next scrape. Loops beat runs.
Read →What Is an AI Systems Architect?
A working definition — from someone who designs, codes, and operates the systems.
Read →I Built an AI Platform Solo — Here's the Architecture
Custom Postgres + RLS, a custom MCP server, and 2,500+ prompts — one operator.
Read →