decimalai-mcp

Part of DecimalAI. Most users want the Python SDK → decimal-labs/decimalai-python.

MCP server for the DecimalAI skills registry — the registry that ranks agent skills by measured effectiveness (verified A/B benchmarks, live pass rates, AI rater scores), not download counts.

Gives any MCP client (Claude Desktop, Claude Code, Cursor, …) three read-only tools:

No API key required — all three tools read public registry endpoints. If you set DECIMAL_API_KEY (from app.decimal.ai/settings), the same tools additionally show which skills your org has already installed (installed_as).

Install

pip install decimalai-mcp
# or, no install needed at config time:
uvx decimalai-mcp

Requires Python 3.10+.

Claude Code

claude mcp add decimalai -- uvx decimalai-mcp
# with an API key:
claude mcp add decimalai -e DECIMAL_API_KEY=dai_sk_... -- uvx decimalai-mcp

Claude Desktop

Add to claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "decimalai": {
      "command": "uvx",
      "args": ["decimalai-mcp"],
      "env": {
        "DECIMAL_API_KEY": "dai_sk_optional"
      }
    }
  }
}

Omit the env block entirely for anonymous read-only access.

Configuration

Why no check_manifest_impact tool?

The manifest-impact endpoint (POST /api/v1/regression-check) is authenticated on the platform — it analyzes your org's production traces against a candidate manifest, so there is no public variant to expose. This server is deliberately a read-only, key-optional public-registry surface. If demand shows up, an authed check_manifest_impact (requiring DECIMAL_API_KEY) is a natural v0.2 addition; the regression-check GitHub Action covers the CI use-case today.

Endpoints used (all public)

• GET /api/v1/registry/skills — browse/search
• GET /api/v1/registry/skills/{slug} — detail
• GET /api/v1/registry/leaderboard — ranked leaderboard (category filtering falls back to the browse endpoint's documented view=ranks mode, because the leaderboard endpoint is uncategorized)

Development

pip install -e ".[dev]"
pytest              # all HTTP mocked; no network
python -m decimalai_mcp.server   # run over stdio

Run pytest yourself before opening a PR. CI also asserts that the pinned mcp<2 still provides FastMCP, which the mocked tests do not cover — run that one too:

python -c "from mcp.server.fastmcp import FastMCP; import decimalai_mcp.server"

Releases are cut from a published GitHub Release — see RELEASING.md for the gates a change has to pass and the version strings that must move together.

License

MIT