Skip to content

MCP Server (deprecated)

Deprecated as of ClawBio 0.7.0, removed in 0.8.0

The MCP server keeps working through the 0.7.x series and prints a notice on stderr when it starts. Nothing in an existing client configuration breaks yet. Plan the move before 0.8.0.

Why

The server shipped in 0.6.0 so that editors other than Claude Code could use the library. Every ClawBio skill is a plain Agent Skills folder, and the editors the server was built for (Cursor, VS Code, Codex, Zed) now read such folders directly from ~/.agents/skills/. A remote-call shim in front of them adds a process, a dependency pin (mcp<2) and a second code path, for no capability the skills do not already have. Removing it leaves one contract to keep correct: the SKILL.md and the script beside it.

Skills that wrap third-party MCP servers (bioqc-mcp, bgpt-mcp, just-prs-mcp) are unaffected. This page is about ClawBio's own server only.

Migrate

Install the plugin. It ships the skills themselves, not a remote-call shim:

/plugin marketplace add ClawBio/ClawBio
/plugin install clawbio

All four read ~/.agents/skills/ (and a project's .agents/skills/). Copy or symlink the skill folders you need there. Each folder is self-contained: SKILL.md, the script, demo data and tests.

git clone https://github.com/ClawBio/ClawBio
mkdir -p ~/.agents/skills
cp -r ClawBio/skills/pharmgx-reporter ClawBio/skills/clinical-variant-reporter ~/.agents/skills/

Use the CLI or the Python API. Both return the same structured result and reproducibility bundle the MCP tools returned.

pip install clawbio
clawbio run pharmgx --demo

from clawbio import run_skill, list_skills
result = run_skill("pharmgx", demo=True)

The three MCP tools map directly: clawbio_list_skills is clawbio list or list_skills(), clawbio_describe_skill is the skill's SKILL.md, and clawbio_run_skill is clawbio run <skill> or run_skill().

Reference for existing configurations

Everything below still works until 0.8.0.

ClawBio ships a Model Context Protocol server, so any MCP-capable client can discover, read, and run bioinformatics skills. It runs locally over stdio. Your data never leaves your machine.

Quick start

No installation required. uvx fetches and runs it in a throwaway environment:

uvx --from 'clawbio[mcp]' clawbio mcp

If you prefer a permanent install:

pip install 'clawbio[mcp]'
clawbio mcp

Client configuration

The same three lines work across clients; only the file location differs.

~/.cursor/mcp.json (global) or .cursor/mcp.json (per project):

{
  "mcpServers": {
    "clawbio": {
      "command": "uvx",
      "args": ["--from", "clawbio[mcp]", "clawbio", "mcp"]
    }
  }
}

~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows:

{
  "mcpServers": {
    "clawbio": {
      "command": "uvx",
      "args": ["--from", "clawbio[mcp]", "clawbio", "mcp"]
    }
  }
}

.vscode/mcp.json in your workspace:

{
  "servers": {
    "clawbio": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "clawbio[mcp]", "clawbio", "mcp"]
    }
  }
}

settings.json:

{
  "context_servers": {
    "clawbio": {
      "command": {
        "path": "uvx",
        "args": ["--from", "clawbio[mcp]", "clawbio", "mcp"]
      }
    }
  }
}

Restart the client, then ask it something like "list the ClawBio pharmacogenomics skills and run the pharmgx demo".

Tools

Tool Purpose
clawbio_list_skills(query) Search the skill catalog. An empty query returns everything. Each result carries a runnable flag.
clawbio_describe_skill(name) Return a skill's full SKILL.md contract: inputs, outputs, safety rules, demo command. Accepts either the skill name or its CLI alias.
clawbio_run_skill(skill, demo, input_path, output_dir, extra_args) Execute a skill and return its structured result, including output file paths.

Not every skill is executable

The catalog contains both executable skills and agent-readable ones that ship only a SKILL.md contract. clawbio_list_skills marks each with a runnable boolean, and clawbio_run_skill refuses non-runnable skills with an explicit message rather than failing obscurely. Agents should not have to guess.

Data safety

By default the server runs demo data only. Passing input_path or output_dir is refused:

This ClawBio MCP server is restricted to demo data. To let it read and write
local files, restart it with CLAWBIO_MCP_ALLOW_LOCAL_FILES=1.

This is deliberate. Adding an MCP server to a client config is a low-friction action, often done once and forgotten, and it should not silently grant an agent read access to a patient genome. To analyse your own files, opt in explicitly:

{
  "mcpServers": {
    "clawbio": {
      "command": "uvx",
      "args": ["--from", "clawbio[mcp]", "clawbio", "mcp"],
      "env": { "CLAWBIO_MCP_ALLOW_LOCAL_FILES": "1" }
    }
  }
}

Skill runs still write a reproducibility/ bundle (commands.sh, environment.yml, SHA-256 checksums) next to their output, so an agent-initiated analysis can be replayed and checked without the original session.

There is no hosted ClawBio MCP endpoint, and this is a design decision rather than a gap. Genomic data should not traverse a third-party server to be analysed.

Requirements

  • Python 3.11+
  • mcp>=1.9,<2 — installed automatically by the [mcp] extra

mcp 2.0

mcp 2.0 removed mcp.server.fastmcp. The clawbio[mcp] extra pins below 2.0. If you installed mcp yourself and see a startup error naming the version, install the extra instead of a bare mcp.

Troubleshooting

Client shows no tools. Confirm the server starts on its own first:

uvx --from 'clawbio[mcp]' clawbio mcp

It should wait silently for input on stdin. Press Ctrl-C to exit. If it exits immediately with a message, the message names the cause.

uvx: command not found. Install uv, or use the pip install 'clawbio[mcp]' route and set "command": "clawbio", "args": ["mcp"].

A skill fails with a missing dependency. Some skills declare extra requirements. Run clawbio_describe_skill and check the dependencies field.