conda-meta-mcp

An MCP (Model Context Protocol) server exposing authoritative, read-only Conda ecosystem metadata for AI agents.

📖 Read the introduction blog post: conda-meta-mcp: Expert Conda Ecosystem Data for AI Agents

What “Meta” Means Here

“Meta” refers to structured, machine-consumable ecosystem intelligence about packages — not the upstream project documentation itself. This server provides (see also the schema server-info.json for current capabilities):

Currently available:

• Version metadata (MCP tool/library versions) via the info tool

• Package info tarball data via the package_insights tool

• Package search via the package_search tool

• Import to package heuristic mapping via the import_mapping tool

• File path to package mapping via the file_path_search tool

• PyPI name to conda package mapping via the pypi_to_conda tool

• CLI help (for conda) via the cli_help tool

• Repository metadata queries (depends / whoneeds) via the repoquery tool

Tools backed by channel-specific data sources require an explicit channel argument and fail for unsupported channels before reading their data source.

Planned:

• Solver feasibility signals (dry-run outputs)

• Schema references and selected spec excerpts

• Binary linkage information

• Links (not copies) to sections of knowledge bases

It does not embed, index, or serve full library docs (e.g. numpy API pages); that remains out of scope by design.

Related MCP server: speclib-mcp

1. Purpose

Enable agents to answer packaging questions by providing up-to-date critical and fragmented expert knowledge. This project provides a safe, inspectable, zero‑side‑effect surface so agents deliver accurate, up‑to‑date guidance.

2. Scope

Goals

• Trustworthy machine interface

• Read‑only, hostable

• Fast startup, low latency

• Clear extension & testing pattern

Non‑Goals

• Performing installs / mutations

• Replacing human docs

• Re‑implementing conda‑forge processing logic

3. Design Principles

• Side‑effect free by contract

• Tool registration pattern (conda_meta_mcp.tools)

• Test + pre‑commit enforced consistency

• Incremental expansion

4. Installation

Via pixi (recommended)

Install globally as a tool:

pixi global install conda-meta-mcp

Or add to your project:

pixi add conda-meta-mcp

Via conda/mamba

conda install -c conda-forge conda-meta-mcp

Or with mamba/micromamba:

mamba install -c conda-forge conda-meta-mcp

From source (development)

Prerequisites: pixi

git clone https://github.com/conda-incubator/conda-meta-mcp.git
cd conda-meta-mcp
pixi run cmm --help

5. Agent Setup

Installed as conda package

Call cmm mcp-json to get an json snippet containing the command with args to add to your agent configuration.

Installed from source

Call pixi run cmm mcp-json to get an json snippet containing the command with args to add to your agent configuration.

6. Usage inside GitHub Copilot coding agent

Create a GitHub workflow named copilot-setup-steps.yml containing (see also GitHub Documentation):

jobs:
  copilot-setup-steps:
    ...
    steps:
      ...
      - name: Setup conda-meta-mcp
        uses: conda-incubator/conda-meta-mcp@main
      ...

Add this MCP configuration inside your repository under Settings -> Copilot -> Coding agent -> MCP Configuration:

{
  "mcpServers": {
    "conda-meta-mcp": {
      "type": "local",
      "command": "cmm",
      "args": [
        "run"
      ],
      "tools": [
        "*"
      ]
    }
  }
}

7. Development

Tasks (pixi):

• Tests: pixi run test (for coverage open htmlcov/index.html)

• Lint / format / type / regenerate metadata: pixi run pre-commit

8. Extending (New Tool)

1. Create conda_meta_mcp/tools/<name>.py with:

   from .registry import register_tool
   
   @register_tool  # or @register_tool(cache_clearers=[...]) for custom cache clearers
   async def my_tool(...) -> dict:
       """Tool description (becomes MCP tool description)."""
       return await asyncio.to_thread(_helper_function, ...)

2. Add unit tests (mock heavy deps)

3. pixi run prek

4. pixi run test

5. Open PR

8. Safety Model

• No environment mutation

• No external command side effects

• Future additions must preserve read‑only contract