pymol-mcp

pymol-mcp

Headless PyMOL as an MCP server — drive molecular visualization, GROMACS/LAMMPS trajectories, and clathrate-hydrate cage science from your LLM.

Demo · Quick Start · Highlights · Features · Tools · Cage Science · 한국어

[!NOTE]An MCP server that embeds PyMOL in-process, headless — no GUI, no socket plugin, no manual setup.It exposes 30+ typed tools, returns rendered images inline so the model can see what it draws,loads GROMACS/LAMMPS trajectories, and ships a clathrate-hydrate analysis toolkit(H-bond networks, F3/F4 order parameters) with numerically validated science.

Demo

Ask in plain language → the model calls typed tools → headless PyMOL renders it. (full-quality MP4)

Highlights

Features

• Embedded & headless — one long-lived PyMOL instance on a dedicated worker thread; nothing to click, works in CI.
• The model can see — render_image ray-traces and returns a PNG as MCP image content.
• MD-native — load GROMACS .gro+.xtc, or bridge LAMMPS/NetCDF/… through MDAnalysis with in-memory coordinate injection.
• Clathrate-hydrate toolkit — H-bond networks and F3/F4 order parameters ported from a validated Rust engine, all in nm with correct triclinic PBC.
• Typed, safe tools — every argument is schema-validated; the arbitrary-code passthrough is opt-in (PYMOL_MCP_ALLOW_CODE_EXEC=1).
• Protocol-hardened — PyMOL's chatty stdout is permanently redirected so it can never corrupt the JSON-RPC stream (with a subprocess test that proves it).

Quick Start

[!IMPORTANT]PyMOL open-source is a conda package, and the server must run in a Python that can import pymol2.Install into that interpreter — do not use uvx/fastmcp install (they build isolated envs without PyMOL).

# 1. Create the environment (or reuse one that already has pymol-open-source)
conda env create -f env.yml        # env named `pymol-mcp`
conda activate pymol-mcp

# 2. Install this package (with the optional MD bridge + dev tools)
pip install -e ".[md,dev]"

# 3. Verify
pytest -q

It's a standard MCP server over stdio, so it works with any MCP-capable client (Claude Code / Desktop,Codex CLI, Gemini CLI, Cline, Continue, …). Point the command at the absolute conda interpreter so itcan import pymol2.

Most clients use an mcpServers block (Claude Code / Desktop, Gemini CLI, Cline, Continue, …):

{
  "mcpServers": {
    "pymol": {
      "command": "/absolute/path/to/conda/envs/pymol-mcp/bin/python",
      "args": ["-m", "pymol_mcp"]
    }
  }
}

[mcp_servers.pymol]
command = "/absolute/path/to/conda/envs/pymol-mcp/bin/python"
args = ["-m", "pymol_mcp"]

Prefer not to hardcode a path? Use "command": "conda", "args": ["run", "-n", "pymol-mcp", "python", "-m", "pymol_mcp"] instead (requires conda on the client's PATH). See llms-install.md for a full from-scratch setup.

To enable the opt-in scripting tools, add "env": {"PYMOL_MCP_ALLOW_CODE_EXEC": "1"} to the server entry.

Then ask your agent things like:

Load ./hydrate.gro, color water by F4 order parameter, and render it.
Load md.gro + traj.xtc, show CO2 guests as spheres, render frame 50.
What's the mean H-bond coordination of the water in this structure?

Tool Catalog

Domain: clathrate-hydrate science

Ported from a validated Rust reference implementation and re-checked against ground truth. All analysis runs innanometres with a correct fractional-coordinate minimum-image convention (orthorhombic and triclinic),a signed atan2 dihedral for F4, and a periodic-image KDTree for neighbour search.

• identify_cages — full TRACE cage perception: ring finding → geometric validation → constraint-propagation assembly → Euler (SEC) validation → face-count typing (5¹², 5¹²6², 5¹²6⁴, …) and an sI/sII/sH structure call.
• cage_occupancy — assign guest molecules (CO₂/CH₄) to cages and report per-type occupancy (θ_S, θ_L).
• mark_cages — drop a colored sphere at each cage centre so render_image can show the cage lattice.
• order_parameter — F4 (torsional) and F3 (three-body angular). F4 ≈ 0.7–0.95 → hydrate, ≈ 0 → liquid, ≈ −0.4 → ice Ih.
• hbond_network — water H-bond graph (O–O ≤ 0.36 nm and a donor H–O···O angle < 35°) with coordination stats.

Detected sII cages drawn as wireframe polyhedra: 5¹² dodecahedra (cyan) around a 5¹²6⁴ cage (red), sharing faces.

[!TIP]Validated against ground truth: on a structure II reference, identify_cages finds exactly **128 × 5¹²

• 64 × 5¹²6⁴** cages (the textbook 2:1 sII lattice), and on structure I exactly 16 × 5¹² + 48 × 5¹²6²;F4 over the first ten waters reproduces the reference value 0.926698 exactly, F3 = 0.0028(hydrate-like ≤ 0.04), and the H-bond network is a perfect tetrahedral (mean coordination 4.00) framework.

How it works

   MCP client (Claude · Codex · Gemini …)
          │  stdio JSON-RPC
          ▼
 ┌───────────────────────────────────────────────┐
 │  pymol-mcp  (FastMCP, conda env with pymol2)    │
 │   • permanent stdout redirect (protocol-safe)   │
 │   • ONE worker thread owns + drives pymol2      │
 │   • typed @mcp.tool functions                   │
 └───────────────────────────────────────────────┘
      │ cmd.* (headless)        │ numpy / scipy (nm)
      ▼                          ▼
  PyMOL 3.x  ── ray → PNG    analysis/ (hbond, F3/F4)
                              coords via iterate_state

Requirements

Contributing

Issues and PRs welcome — see CONTRIBUTING.md.

License

MIT. The optional md extra pulls MDAnalysis (GPL-2.0-or-later), imported lazily; the corepackage stays MIT.