refua-mcp 0.4.0

MCP server for Refua Boltz2 folding/affinity and BoltzGen design workflows.

Refua MCP Server

MCP server exposing Refua's unified Complex API for Boltz2 folding/affinity and BoltzGen design workflows.

Install

pip install refua[cuda] # remove [cuda] if you don't need gpu support
pip install refua-mcp

ADMET predictions are optional; install refua[admet] to enable them:

pip install refua[admet]

Boltz2 and BoltzGen require model/molecule assets. If you don't have them, refua can download them for you automatically:

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

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

MCP Clients

Claude Code

Add the server to your Claude Code MCP config (macOS: ~/Library/Application Support/Claude/claude_code_config.json, Linux: ~/.config/claude/claude_code_config.json). This uses the default assets (~/.boltz for Boltz2 and the bundled BoltzGen artifact). Merge with any existing mcpServers entries:

{
  "mcpServers": {
    "refua-mcp": {
      "command": "python3",
      "args": ["-m", "refua_mcp.server"]
    }
  }
}

Codex

Register the server with the Codex CLI (uses default asset locations):

codex mcp add refua-mcp -- python3 -m refua_mcp.server

List configured servers with:

codex mcp list

If the server is slow to boot (for example on first import of heavy ML deps), raise the startup timeout in your Codex config.toml:

[mcp_servers.refua-mcp]
startup_timeout_sec = 30

Tools

• refua_complex: run a unified Complex spec with action="fold" (default) or action="affinity". Optionally run ADMET for SMILES ligands via admet (auto by default when available).
• refua_job: check status for background jobs and optionally return results.
• refua_admet_profile (optional): run model-based ADMET predictions for SMILES strings (only available when refua[admet] is installed).

Example (fold a protein + ligand with optional affinity):

{
  "tool": "refua_complex",
  "args": {
    "name": "protein_ligand",
    "entities": [
      {"type": "protein", "id": "A", "sequence": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQ"},
      {"type": "ligand", "id": "lig", "smiles": "CCO"}
    ],
    "constraints": [
      {"type": "pocket", "binder": "lig", "contacts": [["A", 5], ["A", 8]]}
    ],
    "affinity": {"binder": "lig"},
    "admet": true
  }
}

Example (ADMET predictions):

{
  "tool": "refua_admet_profile",
  "args": {
    "smiles": "CCO",
    "include_scoring": true
  }
}

Note: DNA/RNA entities are supported for Boltz2 folding only (BoltzGen does not accept DNA/RNA entities).

Long-Running Jobs

For runs that exceed the tool-call timeout, set async_mode=true and poll the job. Folding (action="fold") can take minutes depending on inputs and hardware, so poll sparingly (for example, every 10-30 seconds with backoff) or follow recommended_poll_seconds from refua_job.

{
  "tool": "refua_complex",
  "args": {
    "async_mode": true,
    "entities": [...]
  }
}

Then poll with:

{
  "tool": "refua_job",
  "args": {
    "job_id": "..."
  }
}

For queued/running jobs, the response includes recommended_poll_seconds plus queue and estimate metadata (queue_position = jobs ahead, queue_depth = queued jobs, average_runtime_seconds, estimated_start_seconds, estimated_remaining_seconds) to help clients back off. Set include_result=true once the job is complete to fetch the output.