@dmzagent/skill-builder-mcp

@dmzagent/skill-builder-mcp

An MCP server that connects coding agents — Claude Code, Cursor, Codex, and any other MCP-compatible tool — to the skill-builder registry. Agents can search the registry and auto-install skills (resolving meta-skill dependency trees) directly into the current project.

// One line in your MCP config — no global install needed:
{ "command": "npx", "args": ["-y", "@dmzagent/skill-builder-mcp"] }

Tools

Tool	What it does	Writes files?
skill_search	Search/browse the registry with filters and sorting	No
skill_info	Full details + markdown body for one skill (no download counted)	No
skill_suggest	Autocomplete skills, tags, authors, categories	No
skill_taxonomy	Categories / tags / authors / types with counts	No
skill_install	Install a skill (+ dependencies for meta skills) into the project	Yes

skill_install writes the files each tool expects:

• claude → appends the skill to CLAUDE.md under a ## <name> heading (re-installs are skipped via a hidden <!-- skill-id --> marker).
• codex → same, but to AGENTS.md.
• cursor → .cursor/rules/<slug>.mdc with frontmatter.
• file → <slug>.md + a <slug>.json manifest in output_dir.

With target: "auto" (the default), the server detects the right one from the project (.cursor/, .claude/, CLAUDE.md, AGENTS.md, or env) and falls back to file.

Configuration

All configuration is via environment variables — all optional:

Variable	Default	Purpose
SKILL_API_URL	https://skills.dmzagent.com/api	Registry API base (a bare site URL is fine; /api is appended).
SKILL_TOKEN	(none)	Bearer token; only needed to see your own private/draft skills.
SKILL_PROJECT_DIR	server cwd	Default project root that installs write into.
SKILL_TARGET	(auto-detect)	Force the install target: claude | codex | cursor | file.

Most clients launch the server with the project as its working directory, so installs land in the right place automatically. If yours doesn't, set SKILL_PROJECT_DIR or pass project_dir to skill_install.

Client setup

Claude Code

Add it from the project root, or drop the config into .mcp.json:

claude mcp add skill-builder -- npx -y @dmzagent/skill-builder-mcp

// .mcp.json
{
  "mcpServers": {
    "skill-builder": {
      "command": "npx",
      "args": ["-y", "@dmzagent/skill-builder-mcp"]
    }
  }
}

Cursor

// .cursor/mcp.json (project) or ~/.cursor/mcp.json (global)
{
  "mcpServers": {
    "skill-builder": {
      "command": "npx",
      "args": ["-y", "@dmzagent/skill-builder-mcp"]
    }
  }
}

Codex (OpenAI Codex CLI)

# ~/.codex/config.toml
[mcp_servers.skill-builder]
command = "npx"
args = ["-y", "@dmzagent/skill-builder-mcp"]

Any other MCP client

Launch this command and speak MCP over stdio:

npx -y @dmzagent/skill-builder-mcp

Optionally pass env vars, e.g. SKILL_TARGET=cursor SKILL_PROJECT_DIR=/path/to/project.

Example agent flow

1. skill_search { "query": "literature review", "type": "meta" }
2. skill_info   { "skill": "@kmd_ai/paperdistillery-orchestrated-literature-to-briefing-pipeline" }
3. skill_install{ "skill": "@kmd_ai/paperdistillery-orchestrated-literature-to-briefing-pipeline" }
   → resolves the dependency tree and installs every required skill for the detected tool.

Use dry_run: true on skill_install to preview exactly which files would be written before committing.

Develop / build from source

This package lives in the skill-builder-landing monorepo.

npm install
npm run build --workspace packages/mcp        # tsc → dist/
node packages/mcp/dist/index.js --help        # sanity check
npx @modelcontextprotocol/inspector node packages/mcp/dist/index.js   # interactive testing

The registry data model and types mirror the Worker API; like the other packages in this repo, the MCP server keeps its own copy of the types and the install logic (adapted from packages/cli) so the published package is self-contained.

Listed in the official MCP Registry

This server is published to the official MCP Registry under the domain-based name com.dmzagent/skill-builder-mcp. The registry only stores metadata (in server.json); the package itself lives on npm.

Publishing is automated by .github/workflows/publish-mcp.yml, which runs on a version tag and both publishes to npm and lists in the registry.

One-time setup (DNS authentication)

The com.dmzagent namespace is proven by an apex TXT record on dmzagent.com. Generate an Ed25519 key pair locally — the private half is a CI secret, never committed:

openssl genpkey -algorithm Ed25519 -out mcp-key.pem
# public key (base64) -> terraform var mcp_registry_public_key:
openssl pkey -in mcp-key.pem -pubout -outform DER | tail -c 32 | base64
# private key (hex) -> the MCP_PRIVATE_KEY GitHub secret:
openssl pkey -in mcp-key.pem -noout -text | grep -A3 'priv:' | tail -n +2 | tr -d ' :\n'

Then:

• The public key is committed as the default of mcp_registry_public_key in terraform/mcp-registry.tf, so the deploy.yml terraform apply publishes the TXT record on the next push to main — no local terraform needed. (Override the var in tfvars to rotate the key.)
• Add repo secrets MCP_PRIVATE_KEY (the hex private key) and NPM_TOKEN (npm publish token).

Publish a version

# bump packages/mcp/package.json + server.json to the new version, commit, then:
git tag v1.0.1 && git push origin v1.0.1

To publish manually instead of via CI, from packages/mcp:

npm publish --access public                       # to npm first (registry verifies it)
mcp-publisher login dns --domain dmzagent.com --private-key "$MCP_PRIVATE_KEY"
mcp-publisher publish                              # reads server.json

License

MIT