raspa-mcp

Turn any AI agent into a molecular simulation expert — overnight.

raspa-mcp is a Model Context Protocol (MCP) server that wraps RASPA2 — the gold-standard molecular simulation engine for porous materials — into a clean, agent-friendly tool layer. Feed it a CIF file and a molecule name; let your agent handle the rest.

Why raspa-mcp?

Running RASPA2 correctly requires deep expertise: choosing the right ensemble, setting unit cell replications, picking force fields, validating Ewald summation parameters, and parsing Fortran-style output files. Historically this knowledge lived in the heads of computational chemists and nowhere else.

raspa-mcp encodes that expertise as 20 structured MCP tools — covering every major simulation type RASPA2 supports — so that an LLM agent like featherflow can autonomously design, validate, execute, and interpret molecular simulations without human intervention.

Features at a Glance

Simulation Templates (12 types)

Output Parsing (7 parsers)

• Adsorption loading — mol/kg, mg/g, cm³(STP)/g, molecules/uc, ±errors
• Isosteric heat Qst — from energy fluctuations [kJ/mol]
• Henry coefficient & μ_ex — from Widom insertion, −RT ln(W)
• Helium void fraction — direct extraction
• Radial distribution function g(r) — peak detection, full r/g(r) arrays
• MSD → Diffusion coefficients — self D_s and collective D_c via Einstein relation (NumPy linear fit, latter 50% of trajectory)
• 3D density grid — 2D slice extraction from .grid files
• Thermodynamic Integration — trapezoidal ∫⟨∂U/∂λ⟩dλ → ΔA [kJ/mol]
• Multi-component mixture — per-component loading with backward compatibility

Analysis Tools

• Selectivity S_AB — (x_A/x_B) / (y_A/y_B) from mixture loadings
• Isotherm plotting — single and multi-MOF comparison PNGs (matplotlib)
• Density slice plotting — heatmap PNG from 3D grid data

Built-in Knowledge Base

• 6 molecules: CO2, N2, CH4, H2O, helium, n-butane (TraPPE / SPC-E)
• 5 force fields: TraPPE-CO2/N2/CH4/H2O, UFF — with mixing rules, pseudo-atom definitions
• Input validator — catches 20+ common mistakes before RASPA2 ever runs
• Environment checker — reports RASPA2 readiness on server startup

Installation

0. Install uv (if not already present)

# Linux / macOS
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

uv is a fast Python package manager. It replaces pip + venv with a single tool and handles the Python version automatically.

1. Clone and install the Python package

git clone https://github.com/lichman0405/raspa-mcp
cd raspa-mcp
uv sync

uv sync creates a virtual environment under .venv/, pins the Python version (3.11+), and installs all dependencies — no manual venv or pip needed.

2. Compile and configure RASPA2 (one-time setup)

uv run python -m raspa_mcp.installer

This single command:

• Checks for build tools (git, gcc, make, autoconf, automake, libtool) and installs any that are missing automatically via apt-get / dnf / yum / brew etc.
• Clones RASPA2 from GitHub and compiles from source
• Installs to ~/.local/raspa2 by default (override with --prefix)
• Writes RASPA_DIR and PATH exports to your shell RC file automatically

# custom install prefix:
uv run python -m raspa_mcp.installer --prefix /opt/raspa2

# force reinstall even if already present:
uv run python -m raspa_mcp.installer --force

After the command finishes, reload your shell:

source ~/.bashrc   # or ~/.zshrc, ~/.profile, etc.

Compatibility: Tested on Ubuntu 24.04 (GCC 14). The installer automatically applies -std=gnu11 and other compiler flags to work around known issues in the RASPA2 source tree on modern GCC versions.

RASPA2 source: https://github.com/iRASPA/RASPA2 Reference: D. Dubbeldam, S. Calero, D.E. Ellis, R.Q. Snurr, Mol. Simul. 42, 81–101 (2016)

Quickstart — featherflow

Register raspa-mcp with one command (run in the featherflow project directory):

featherflow config mcp add raspa2 \
  --command uv \
  --arg run \
  --arg --directory \
  --arg /path/to/raspa-mcp \
  --arg raspa-mcp \
  --lazy \
  --description "RASPA2 molecular simulation: GCMC, MD, adsorption isotherms, force fields, output parsing" \
  --timeout 600

--lazy: raspa-mcp exposes 20 tools — lazy mode registers a single gateway entry-point instead of all tools upfront, keeping per-call LLM token cost low. --timeout 600: RASPA2 simulations can take minutes; featherflow recommends 300–600 s for scientific computing MCP servers.

Verify:

featherflow config mcp list

Your agent can now autonomously:

User: Study CO2 adsorption in ZIF-8 at 298 K from 0.1 to 50 bar.

Agent:
  1. raspa-mcp.get_simulation_template("VoidFraction")   → run RASPA2
  2. raspa-mcp.parse_raspa_output(...)                   → void fraction = 0.47
  3. raspa-mcp.get_simulation_template("GCMC")           → fill placeholders × 7 pressures
  4. raspa-mcp.validate_simulation_input(...)            → clean
  5. shell_exec → RASPA2 × 7
  6. raspa-mcp.parse_raspa_output(...)                   → isotherm data
  7. raspa-mcp.plot_isotherm(...)                        → ZIF-8_CO2.png
  8. feishu-mcp.upload_file_and_share(...)               → report delivered

No human intervention required.

MCP Tools Reference

Custom-everything workflow

When you bring your own CIF, your own force field, and your own molecule definitions, the tooling above gives you safe builders for every file RASPA2 expects:

inspect_cif(cif)               → formula, cell, charges, overlap warnings
recommend_supercell(cif, 12)   → UnitCells line + ChargeMethod hint
create_workspace(work, name, cif)
generate_force_field_def(work)                       # safe "3 zeros" overwrite file
generate_force_field_mixing_rules_def(work, atom_types=[...])  # LJ ε/σ
generate_pseudo_atoms_def(work, atoms=[...])         # atom registry
generate_molecule_def(work, "MyAdsorbate", ...)      # per-molecule .def
preflight_workspace(work)                            # cross-file sanity

`force_field.def` ≠ `force_field_mixing_rules.def`. The first file is overwrite rules (use the "3 zeros" minimum for almost every job). The second file is where Lennard-Jones ε/σ live. Mixing them up is the single most common cause of cryptic :# parse errors from RASPA2. The generators emit the correct format for both.

For an end-to-end recipe call get_workflow_recipe("custom_mof_gcmc").

All write operations are sandboxed under RASPA_MCP_WORKSPACE_BASE (default ~/raspa_workspaces) — paths outside that root are rejected.

Testing

uv run pytest tests/ -q          # 41 tests, ~1.5 s
uv run ruff check raspa_mcp/ tests/

Architecture

raspa-mcp/
├── raspa_mcp/
│   ├── server.py        # 20 MCP tools (FastMCP, stdio transport)
│   ├── parser.py        # Output parsers (loading, RDF, MSD, TI, density)
│   ├── validator.py     # Input validator (20+ rule checks)
│   ├── installer.py     # RASPA2 env detection + source build + raspa-mcp-setup CLI
│   └── data/
│       ├── templates.py    # 12 simulation.input templates
│       ├── molecules.py    # 6 molecule definitions + metadata
│       └── forcefields.py  # 5 force field file sets
├── tests/
│   └── test_server.py   # 41 unit tests
└── docs/
    └── workflow.md      # Full autonomous research workflow walkthrough

Full Business Workflow

See docs/workflow.md for a complete end-to-end walkthrough of an autonomous MOF screening study using featherflow + raspa-mcp + RASPA2 + feishu-mcp, from a single chat message to a ranked report delivered to Feishu — approximately 120–140 tool calls, zero human steps.

License

MIT

Acknowledgements

Built on top of RASPA2 by Dubbeldam, Calero, Ellis & Snurr. Force-field parameters from the TraPPE family (Martin, Siepmann et al.) and the Universal Force Field (Rappé et al.).