milo-mcp

MCP server for the Milo/EDS ecosystem. Gives AI assistants (Claude, Cursor) a context-aware interface to browse, search, and inspect Milo blocks (write/scaffold tools are planned for later phases).

Phase 1 — Foundation (current)

Read-only tools for block discovery and inspection:

Tool	Description
get_block_repo_owner	Project key + GitHub org/repo for the resolved block (same resolver as get_block_location)
get_block_codeowner	CODEOWNERS teams/individuals (Milo when block exists in core) + commit activity on the block directory
get_block_location	File path + GitHub blob URL for the block’s main .js
get_block	Fetch block metadata + optional source/CSS/tests
search_blocks	RAG-powered discovery (Adaptive RAG routing)
list_blocks	List blocks; optional include_child_overrides compares child overrides to Milo core and reports lag
get_block_history	Git commit history for a block
get_preview_url	AEM URLs: env is stage (default, *.aem.page) or live (*.aem.live); also returns milo_override_url and local_url
refresh_index	Rebuild the block index (one project or all) without restarting the server
analyze_block_dependencies	Inspect which blocks a block imports, and optionally the reverse direction (expensive)
get_override_migration_guide	For a child-project override, show what changed in Milo core since it was last updated; optional LLM-generated migration steps
add_project	Register a new GitHub repo at runtime (custom projects persist across restarts)
remove_project	Remove a custom project; cannot remove built-in defaults
list_projects	List registered projects with source (default/custom/custom_override) and indexed block count
check_setup_status	Diagnose setup: verify GITHUB_TOKEN / ANTHROPIC_API_KEY with live test calls, report index state and per-project block counts

Resources

• milo://conventions/block-anatomy — canonical JS/CSS/test structure
• milo://conventions/da-table-syntax — DA markdown block tables
• milo://conventions/nala-test-pattern — Playwright test structure
• milo://conventions/codeowners — where Milo’s CODEOWNERS lives and how tools use it
• milo://project/{project}/fstab — child project fstab.yaml (e.g. da-bacom, bacom, cc are listed in MCP; any known project can be read by URI)

Prompts

• audit-child-project-blocks — guided audit of stale overrides vs Milo core (uses list_blocks with include_child_overrides=true)

Install

Two distribution channels. Pick whichever fits your client.

Claude Desktop — one-click bundle

Download the latest *.mcpb file from the Releases page and double-click it. Claude Desktop opens an install dialog; fill in your GITHUB_TOKEN and optionally ANTHROPIC_API_KEY, then click Install. No local Node install needed — Claude Desktop ships its own runtime and the embedding model is pre-bundled in the file.

To uninstall, open Claude Desktop → Settings → Extensions, find milo-mcp, and click Remove.

Any MCP client — npm package

Published as @addumone/milo-mcp on the public npm registry. Configure your MCP client to invoke it via npx:

{
  "mcpServers": {
    "milo": {
      "command": "npx",
      "args": ["-y", "@addumone/milo-mcp"],
      "env": {
        "GITHUB_TOKEN": "ghp_...",
        "ANTHROPIC_API_KEY": "sk-ant-..."  // optional — enables CRAG; omit for cosine-only search
      }
    }
  }
}

No .npmrc setup needed — the package is on the public registry. The embedding model is downloaded from Hugging Face on first run and cached locally; subsequent starts are instant. To force the latest version: npx --prefer-online @addumone/milo-mcp.

Local development

cp .env.example .env
# GITHUB_TOKEN is required; ANTHROPIC_API_KEY is optional (enables CRAG for search_blocks)
npm install
npm run dev

The server loads .env from the process working directory (dotenv) when you run npm run dev or npm start; set the same vars in your MCP host config if the cwd has no .env.

Commands

npm run dev              # Development mode with hot reload
npm test                 # Run unit tests (Vitest)
npm run test:integration # Integration tests under test/integration/ (separate Vitest config)
npm run eval             # Run RAGAS evaluation (Phase 1 gate)
npm run build            # Compile TypeScript and copy convention markdown into dist/
npm start                # Run compiled server

RAG Strategy

Queries are classified into four types and routed to the most appropriate technique:

Type	Technique	Example
LOOKUP	Direct fetch	"get the accordion block"
SEMANTIC	Cosine + CRAG (1–2 word queries try a direct name match first)	"find a block for promotional content"
COMPOSITIONAL	Iterative (3-pass)	"which blocks override milo and are stale?"
MULTI_SOURCE	Semantic + CRAG (same path as SEMANTIC today; agentic multi-tool flow is Phase 4)	Queries mentioning Figma, DA, or “new page”

Environment Variables

See .env.example for all variables.

Required:

• GITHUB_TOKEN — read:repo scope (indexing + GitHub-backed tools)

Optional:

• ANTHROPIC_API_KEY — enables CRAG semantic validation in search_blocks. Without it, search_blocks still works but uses cosine-similarity ranking only, which produces lower-quality results. All other tools are unaffected. A warning is logged at startup when this key is not set.
• Optional tuning (see .env.example; defaults in src/config.ts): MILO_REPO_*, EMBEDDING_MODEL, CRAG_THRESHOLD, SELF_RAG_MAX_ATTEMPTS, COSINE_GAP (CRAG candidate band vs top cosine score, default 0.20).

Required for write tools (Phase 3+):

• GITHUB_WRITE_TOKEN — repo scope

Future phases (documented in .env.example): FIGMA_API_KEY, DA_API_TOKEN.