Connect via MCP

commonspecs speaks the Model Context Protocol — one remote HTTP endpoint that exposes the product lookup as native tools. The reference skill is still the primary, day-one path; MCP is for clients that prefer a native tool connection.

Endpoint

https://api.commonspecs.com/mcp

Phase 1 is bearer-authenticated: send your cs_live_… token as Authorization: Bearer … (the same token the skill uses). Browser-based OAuth login is on the roadmap. Transport: stateless Streamable HTTP, MCP spec 2025-11-25.

1. Add the server

Claude Code:

claude mcp add --transport http commonspecs https://api.commonspecs.com/mcp \
  --header "Authorization: Bearer cs_live_…"

Other clients (Cursor, VS Code, Claude Desktop connectors) take an HTTP MCP server config with the same URL and header. For a stdio-only client, bridge it:

npx mcp-remote https://api.commonspecs.com/mcp \
  --header "Authorization: Bearer cs_live_…"

2. Tools

Four read tools, each a thin wrapper over the REST API — same data, same per-field confidence:

• lookup_product — resolve one product by url, ean, or brand+model → specs + per-field confidence
• search_products — discover products in a category, ranked by quality score
• compare_products — 2–5 products side by side + the overall quality winner
• get_rankings — top-N in a category

Results carry per-field confidence and a quality_score; the scoring weights are intentionally not exposed. Contributing values back over MCP (a write tool) comes in a later phase — until then, the skill is the contribution path.

3. Resources

Read-only context your client can attach once, instead of spending a tool call:

• commonspecs://categories — every category slug
• commonspecs://category/{slug}/schema — the public field names and units for a category (never the scoring weights)
• commonspecs://methodology — how the quality score is approached, in plain language

See the API reference for the underlying endpoints, or the methodology for how the score is built.