vedette-evidence-mcp
A small MCP (Model Context Protocol) server, built to close a specific, honest gap: hands-on experience authoring an MCP server with real tools.
What it is
A working MCP server exposing three tools over stdio:
- search_evidence — keyword search over a seeded set of evidence chunks, each traceable to a named source document.
- list_opportunities — lists opportunities, each linked to the evidence chunk IDs that support it.
- compute_roi — deterministic ROI math (hours saved, adoption ramp, annual dollars saved). The model may propose the inputs; every dollar figure is computed in plain TypeScript, never generated by a model. This mirrors the same model/computation separation used in Vedette.
What it is honestly
This is a demo project with seeded, fictional data (src/data.ts), built to
learn and demonstrate the MCP authoring pattern: defining tools with Zod
schemas, running a stdio transport, and having a client actually call them.
It is not connected to a live database or a real Finance system.
Verified end to end: test-client.mjs connects a real MCP client to the
built server, lists the tools, and calls all three — output captured in
git history / can be re-run any time with node test-client.mjs.
Honest resume line
"Built and tested an MCP server exposing evidence search, opportunity listing, and deterministic ROI computation as callable tools, verified end-to-end with a real MCP client."
Do not claim: connection to a live production system, multi-agent orchestration (this is tool-calling by a single client, not multiple agents), or use by real users. This is a learning/demo artifact, described as exactly that.
Running it yourself
npm install
npx tsc
node test-client.mjs
Next step, if you want to extend it
Swap src/data.ts for a real Supabase query against Vedette's existing
schema (evidence chunks and opportunities tables already exist there) —
that would upgrade this from a demo to a real integration, and would be an
honest next resume line at that point, not before.