@upstate-web/uwc-pipeline-mcp

MCP server that exposes a directory of Claude Agent Skills as callable tools — list_skills and get_skill. Pairs with @upstate-web/uwc-skills, but works against any Agent Skills-compliant directory.

What this is

A minimal Model Context Protocol server that surfaces Agent Skills over stdio. Point it at a directory of SKILL.md-compliant subdirectories and Claude Code (or any MCP-compatible client) can list and fetch them as structured tool calls.

Two tools shipped in v0.1.0:

| Tool | Arguments | Returns | |---|---|---| | list_skills | none | Slug, version, name, and description for every skill in the configured directory | | get_skill | slug: string | Full SKILL.md content (frontmatter + body) for the named skill |

The server is deliberately content-agnostic. The bundled @upstate-web/uwc-skills is one valid skill source; anything that follows Anthropic's Agent Skills spec will work.

Install

npm install -g @upstate-web/uwc-pipeline-mcp

Or install locally and invoke via npx:

npx @upstate-web/uwc-pipeline-mcp

Configure your skills directory

The server reads skills from the first match:

1. UWC_SKILLS_DIR environment variable, if set
2. ./node_modules/@upstate-web/uwc-skills/skills, if the package is installed in the working directory

If neither exists, list_skills returns an empty list with instructions.

# Point at an arbitrary directory
export UWC_SKILLS_DIR=/path/to/my/skills
npx @upstate-web/uwc-pipeline-mcp

Use with Claude Code

Register as an MCP server:

claude mcp add uwc-pipeline-mcp -- npx @upstate-web/uwc-pipeline-mcp

With a custom skills directory:

claude mcp add uwc-pipeline-mcp \
  --env UWC_SKILLS_DIR=/path/to/skills \
  -- npx @upstate-web/uwc-pipeline-mcp

Then in a Claude Code session, the tools appear as list_skills and get_skill. A typical flow:

1. Claude calls list_skills → sees all available skill slugs + descriptions
2. Based on the current task, Claude calls get_skill(slug) to load the full skill content
3. Claude applies the skill's guidance to your code

Smoke test

Verify the server works from the command line by piping JSON-RPC over stdio:

echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized"}
{"jsonrpc":"2.0","id":2,"method":"tools/list"}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"list_skills","arguments":{}}}' | UWC_SKILLS_DIR=/path/to/skills npx @upstate-web/uwc-pipeline-mcp

Scope & non-goals

This package ships a deliberately narrow surface area:

• Read-only. No write tools; no mutation of skill content.
• No network calls. Reads local filesystem only.
• No pipeline-specific tools. generate_phase_prompt, run_governance_agent, run_meta_retro_prompt, compile_knowledge_bundle are not shipped and not on the roadmap. Each requires UWC-specific schema, endpoints, or directory layout that would bind this package to a private stack. They fail our leakage test. Dropped, not deferred.
• No authentication. stdio is the only transport; auth is the MCP client's responsibility.

If you need pipeline tooling beyond skill browsing, the source of truth for those patterns is published as skill content in @upstate-web/uwc-skills — load the relevant skill with get_skill and apply it yourself.

New tools enter this server only if they pass the same bar as new skills: portability (works without UWC infrastructure), leakage-free (no private schema, keys, or endpoints), and usefulness (value to a non-UWC developer). list_skills + get_skill are both read-only filesystem queries against a configurable directory — the canonical shape we'll stick to.

Build from source

git clone https://github.com/upstate-web-co/uwc-pipeline-mcp
cd uwc-pipeline-mcp
npm install
npm run build
node dist/index.js

Related

• Skill bundle: @upstate-web/uwc-skills
• Agent Skills spec: docs.anthropic.com/en/docs/agents/skills
• MCP spec: modelcontextprotocol.io
• MCP TypeScript SDK: @modelcontextprotocol/sdk

License

MIT. Use freely in commercial work.