Refua

Fold and score biomolecular complexes and optionally profile ADMET to prioritize molecules in drug discovery pipelines via the refua-mcp server.

Security Scan

The skill's instructions, requirements, and behavior are internally consistent with a Refua MCP integration for running folding/ADMET workloads; no unexplained credential or install demands were found.

Purpose & Capability

The SKILL.md describes a Refua-to-MCP integration (folding, affinity scoring, optional ADMET) and all declared requirements match that purpose: it requires running a refua-mcp server and optionally installing refua/refua-mcp and model assets. There are no unrelated environment variables, binaries, or config paths requested.

Instruction Scope

Runtime instructions are focused: pip-install Refua/refua-mcp, download model assets, and run the MCP server (python3 -m refua_mcp.server). The instructions do not ask the agent to read unrelated files or exfiltrate data. Note: starting the MCP server exposes a network service (local or remote) — the operator should consider access controls and firewalling when launching it.

Install Mechanism

This is an instruction-only skill (no install spec). It tells an operator to pip install packages from PyPI and to run a Python asset-download helper. Using pip is expected for a Python ML tool, but pip installs execute arbitrary code from the package; operators should install into an isolated virtualenv/container and verify the package source (the SKILL.md links to the GitHub repo). Model asset downloads may be large and come from external hosts (e.g., HF/GitHub), so check bandwidth, integrity, and provenance.

Credentials

The skill declares no required environment variables, credentials, or config paths. That is proportionate to the stated goal (local MCP server + local model assets).

Persistence & Privilege

The skill does not request always:true, does not request to modify other skills, and contains no install-time operations that would permanently alter agent-wide configuration. Agent autonomous invocation remains allowed (platform default) but is not combined with other concerning privileges.

Assessment

This skill appears coherent for running Refua via an MCP server, but take basic operational precautions before installing/running it: - Install into an isolated Python environment or container (pip packages run code at install time). - Verify the package source/repository (the SKILL.md points to github.com/agentcures/refua-mcp) and review maintainers if you need a higher trust level. - Expect large model downloads and heavy CPU/GPU usage; check disk, GPU drivers, and bandwidth limits before downloading assets. - When running the MCP server, restrict network exposure (bind to localhost, use firewall rules, or run in an internal network) so the service isn't unintentionally reachable from the public internet. - No credentials are requested by the skill, but if you add auth or remote endpoints later, ensure secrets are scoped and stored securely. - Do not use results as clinical or regulatory guidance without domain expert review; the SKILL.md itself warns against wet‑lab/clinical guidance. If you want a deeper check, provide the refua/refua-mcp PyPI package names or the repository contents so I can review code-level behaviors (install-time scripts, remote URLs used for downloads, and any network/server authorization defaults).

Skill: Refua

Summary

Refua is used in drug discovery to computationally fold and score biomolecular complexes (e.g., protein–ligand/protein–protein) and optionally profile ADMET, helping prioritize which molecules to synthesize and test first in a drug discovery pipeline.

This skill runs and connects to the refua-mcp MCP server, which exposes Refua’s “unified Complex API” as MCP tools for:

• Boltz2 complex folding (+ optional affinity evaluation)
• BoltzGen design workflows
• Optional ADMET profiling (when installed)

Clawdbot supports MCP natively, so the only requirement is running this MCP server and calling its tools. (github.com)

---

When to use

Use this skill when you need to:

• Fold a protein–ligand, protein–protein, or (fold-only) DNA/RNA complex
• Estimate binding affinity for a specified binder within a complex spec
• Run ADMET predictions for one or more SMILES ligands (if enabled)
• Execute GPU/CPU-heavy Refua workflows via MCP tool calls

Do NOT use this skill when:

• The task is a simple deterministic calculation (prefer a non-ML tool)
• The user expects you to invent sequences/SMILES (request inputs instead)
• The user requests unsafe wet-lab or clinical guidance

---

Installation & assets (operator steps)

1) Install Refua + refua-mcp

Install Refua (CPU or CUDA), then install the MCP server package: (github.com)

• GPU support:
  • pip install refua[cuda]
• CPU-only:
  • pip install refua
• MCP server:
  • pip install refua-mcp

2) Optional: enable ADMET

ADMET tool support is optional and requires an extra: (github.com)

• pip install refua[admet]

3) Download model/assets

Boltz2 and BoltzGen require model/molecule assets. Refua can download them automatically: (github.com)

• python -c "from refua import download_assets; download_assets()"

Default asset locations + overrides: (github.com)

• Boltz2 uses ~/.boltz by default
  • Override via tool option boltz.cache_dir if needed
• BoltzGen uses a bundled HF artifact by default
  • Override via tool option boltzgen.mol_dir if needed

---

Running the MCP server

Start the server using the module entrypoint: (github.com)

python3 -m refua_mcp.server