mcp-analytics-stub
# mcp-analytics-stub
A minimal [Model Context Protocol](https://modelcontextprotocol.io) server in TypeScript. It exposes two tools over stdio and returns **synthetic data only**. It exists to show the shape of an MCP server that wraps an analytics API, without wrapping a real one.
## Tools
| Tool | Kind | What it does |
|------|------|--------------|
| `resolve_category` | resolver | Turns a name fragment into stable category ids. |
| `get_insights` | context | Given an id, returns a few insight rows. Unknown ids return an error result. |
## Shape
```
MCP client (IDE agent)
│ JSON-RPC over stdio
▼
MCP server ── zod validation ──▶ resolver / context handlers ──▶ data layer
```
In a real server the data layer would be an HTTPS client, with credentials read from environment variables. Here it is an in-memory list.
## Why these two tool kinds
- **Resolvers** map human names to ids, so the agent never has to guess an identifier.
- **Context tools** take an id and return structured data, so answers come from a call rather than from pasted text.
- Inputs are schema-validated, so a malformed call fails loudly.
- Tool descriptions matter: they are what an agent reads when choosing a tool.
## Run it
```bash
npm install
npm test
npm run dev # stdio server
```
To use it from an MCP-capable editor, point it at `npx tsx src/server.ts` (or `node dist/server.js` after `npm run build`) as a stdio server.
## From local to shared
The same handlers can sit behind a stateless HTTP transport in a container for team use. This stub keeps stdio only, to stay small.
## License
MIT. All data is invented.
TDQS
Scored across 2 tools
resolve_category and get_insights target clearly different stages of the workflow (discovery vs. retrieval), and the descriptions explicitly explain the handoff via category ids. There is no overlap that could cause misselection.
Both tools follow a consistent verb_noun snake_case pattern. The convention is applied uniformly with no mixed styles or casing deviations.
Two tools is thin but the stub scope (resolve a category, fetch its insights) is plausibly minimal. It sits at the borderline where an agent may need more surface for real work.
Coverage is a narrow two-step path: discovery requires a query string with no way to list all categories, and there is no pagination, time range, or drill-down beyond 'top insights'. Core workflow works but obvious gaps remain.