LLM Skills POC

This monorepo packages one shared pr-review skill and exposes it through both an HTTP API and an MCP server. The skill manifest, instructions, examples, and audit traces are shared across both entrypoints so API clients and MCP clients run the same contract.

Stack

• Node.js 22

• TypeScript

• npm workspaces

• Express HTTP API

• Model Context Protocol server over stdio

• Zod and YAML for skill contracts and manifests

• Vitest for tests

• Docker Compose for the API container

Related MCP server: AI-PR-Review-MCP

Install, Test, And Build

npm install
npm test
npm run build

Local API

Start the API in development mode:

npm run dev:api

After npm run build, start the built API:

npm run start:api

Health check:

curl http://localhost:3000/healthz

List skills:

curl http://localhost:3000/skills

Run the shared pr-review skill:

curl -X POST http://localhost:3000/skills/pr-review/run \
  -H 'Content-Type: application/json' \
  -d '{
    "client": "api",
    "inputs": {
      "repo": "checkout-service",
      "diff": "diff --git a/src/cart.ts b/src/cart.ts\n+export function formatCart(items) {\n+  return items.map(item => item.name).join(\", \");\n+}",
      "risk_level": "low"
    }
  }'

Admin Workbench

Start the API:

npm run dev:api

Open:

http://localhost:3000/admin

The workbench can load skills, edit skill.yaml, edit instructions.md, edit packaged examples, save validated changes, restore the latest backup, run a skill, evaluate examples, and inspect redacted audit events.

Every write creates a timestamped backup under .data/backups/<skill-id>/.

Local MCP

Run the MCP server in development mode:

npm run dev:mcp

After npm run build, MCP clients can run the built stdio server with:

node apps/mcp-server/dist/index.js

Available MCP tools:

• list_skills

• get_skill

• run_skill

• evaluate_skill

Docker

Build and run the API service:

docker compose up --build

The compose service publishes the API on port 3000, mounts ./skills read-write for admin edits, and mounts ./.data for traces and backups.

Check the containerized API:

curl http://localhost:3000/healthz

Build the stable API image, then run the MCP server from that image:

docker compose build
docker run --rm -i llm-skills-poc-api node apps/mcp-server/dist/index.js