Rosetta MCP Server

Author: Ariel J. Ben-Sasson

A Model Context Protocol (MCP) server that lets Cursor (or any MCP client) work with Rosetta, PyRosetta, and Biotite: run RosettaScripts, validate XML protocols, translate between Rosetta and Biotite, score structures, and query documentation -- all from your AI coding assistant.

What's new in v1.3.0 (vs v1.1.8 on npm)

New: Biotite integration

• rosetta_to_biotite -- Find the Biotite equivalent of any Rosetta function with working example code (21 mappings covering structure I/O, SASA, RMSD, superimposition, secondary structure, contacts, hydrogen bonds, B-factors, angles, and more)
• biotite_to_rosetta -- Reverse lookup: find the Rosetta equivalent of a Biotite function
• translate_rosetta_script_to_biotite -- Translate entire RosettaScripts XML or PyRosetta code to Biotite Python. Design/optimization operations are flagged as Rosetta-only.
• Fuzzy search with keyword aliases ("contacts", "binding energy", "surface area", "align", etc.)

Improved: XML to PyRosetta translator

• 37 element types supported (was 6): 11 movers, 9 filters, 10 selectors, 7 task operations
• Full attribute handling: repeats, disable_design, cartesian, tolerance, threshold, distance, and more
• Child element support: MoveMap (with Span), Reweight, ScoreFunction
• Reports unrecognized elements so you know what needs manual work

Improved: Help and documentation

• get_rosetta_help now accepts any topic: movers by name ("FastRelax"), concepts ("constraints", "docking"), or score functions ("ref2015") -- auto-fetches live docs from rosettacommons.org
• search_rosetta_web_docs fallback: when DuckDuckGo is rate-limited, probes direct Rosetta docs URLs
• get_cached_docs auto-caches: no need to call cache_cli_docs first
• Expanded static help for score_functions, movers, filters, xml, and parameters

Improved: Scoring

• pyrosetta_score: new per_residue option returns per-residue energy breakdown
• scorefxn parameter now works (was ignored in v1.1.8)
• Proper error messages for missing files instead of silent {}

Improved: Validation

• validate_xml: new validate_against_schema option checks element names against the Rosetta XSD schema (catches typos like FastRleax)

MCP spec compliance fixes

• tools/call responses now use correct { content: [{ type: "text", text }] } format
• Tool errors return isError: true (not JSON-RPC errors)
• Standard JSON-RPC error codes (-32601, -32700, -32603)
• Removed false resources capability advertisement

Security fixes

• User input no longer interpolated into Python code (uses env vars / stdin)
• Temp files written to os.tmpdir() (not module directory)

Cleanup

• Removed 3 redundant tools: list_functions (merged into get_rosetta_info), search_pyrosetta_wheels, cache_cli_docs (auto-cache in get_cached_docs)
• Removed hardcoded personal paths
• Fixed shadowed variables, async anti-patterns, dead code
• 18 tools (was 21), all with improved agent-oriented descriptions

Example: asking a naive question

This is what makes the MCP server powerful -- an AI agent can answer domain questions by calling the right tools automatically:

User asks in Cursor: "How do I relax my protein and what's the Biotite equivalent?"

The agent calls two MCP tools behind the scenes:

1. get_rosetta_help("FastRelax") returns 6000+ chars of live documentation:

FastRelax performs all-atom relaxation using the FastRelax protocol. Parameters include scorefxn, repeats, cartesian, disable_design, MoveMap configuration...

2. rosetta_to_biotite("FastRelax") returns:

{
  "found": true,
  "results": [{
    "rosetta": { "name": "FastRelax", "example": ["relax = FastRelax()", "relax.set_scorefxn(get_score_function('ref2015'))", "relax.apply(pose)"] },
    "biotite": null,
    "equivalence": "none_from_biotite",
    "notes": "Biotite does NOT perform structure optimization. These are Rosetta-specific capabilities."
  }]
}

The agent synthesizes: "FastRelax is Rosetta's all-atom relaxation protocol. Here's how to use it... Note: Biotite is analysis-only and has no equivalent -- you need PyRosetta for structure optimization."

Without the MCP, the agent would guess from training data and likely get parameter names or API signatures wrong.

What you get (18 tools)

Discovery & Help

Documentation

Execution & Scoring

Translation

Validation & Schema

Environment

Quick start

1. Install from npm

npm install -g rosetta-mcp-server

2. Set up Python environment

# Create a venv with PyRosetta and Biotite
uv venv ~/.venvs/rosetta-mcp
~/.venvs/rosetta-mcp/bin/pip install pyrosetta-installer biotite
~/.venvs/rosetta-mcp/bin/python -c "import pyrosetta_installer as I; I.install_pyrosetta()"

Or skip this step -- PyRosetta auto-installs on first use (takes 10-30 min).

3. Configure your MCP client

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "rosetta": {
      "command": "rosetta-mcp-server",
      "args": [],
      "env": {
        "ROSETTA_BIN": "/path/to/rosetta_scripts.default.macosclangrelease",
        "PYTHON_BIN": "/path/to/.venvs/rosetta-mcp/bin/python"
      }
    }
  }
}

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "rosetta": {
      "command": "rosetta-mcp-server",
      "env": {
        "ROSETTA_BIN": "/path/to/rosetta_scripts.default.macosclangrelease",
        "PYTHON_BIN": "/path/to/.venvs/rosetta-mcp/bin/python"
      }
    }
  }
}

Environment variables:

4. Restart your editor

Open Settings -> MCP. The "rosetta" server should appear green with 18 tools.

XML to PyRosetta translation example

Input XML:

<ROSETTASCRIPTS>
  <SCOREFXNS>
    <ScoreFunction name="ref" weights="ref2015"/>
  </SCOREFXNS>
  <RESIDUE_SELECTORS>
    <Chain name="chainA" chains="A"/>
  </RESIDUE_SELECTORS>
  <MOVERS>
    <FastRelax name="relax" scorefxn="ref" repeats="5" cartesian="true"/>
  </MOVERS>
  <PROTOCOLS>
    <Add mover="relax"/>
  </PROTOCOLS>
</ROSETTASCRIPTS>

Generated PyRosetta code:

import pyrosetta
from pyrosetta import pose_from_pdb
from pyrosetta.rosetta.core.scoring import get_score_function
from pyrosetta.rosetta.core.select.residue_selector import *
from pyrosetta.rosetta.protocols.relax import *

pyrosetta.init("-mute all")

pose = pose_from_pdb("your_protein.pdb")

# Residue Selectors
chainSelector = ChainSelector()
chainSelector.set_chain_strings("A")

# Movers
fastRelax = FastRelax()
fastRelax.set_scorefxn(get_score_function("ref"))
fastRelax.set_default_repeats(5)
fastRelax.cartesian(True)

sfxn = get_score_function("ref2015")

# Apply movers
fastRelax.apply(pose)

pose.dump_pdb("output.pdb")
score = pose.energies().total_energy()
print(f"Final score: {score}")

Rosetta <-> Biotite mapping coverage