pocketscout-mcp: Instructions file for Claude Code

PocketScout MCP

What This Is

PocketScout is a fast triage tool for drug-target binding sites — it gives an AI assistant the tools to pull together everything known about a protein's pockets (structure, chemistry, conservation, literature) into a single briefing in minutes. The flagship use case is reconnaissance before computational binder design: filling the gap between "I have a target" and "I'm running RFdiffusion."

Built with FastMCP 3.0, Python 3.11+, httpx, Pydantic v2.

Who This Is For

Drug discovery scientists (computational and bench) who need a fast binding-site briefing — whether evaluating an unfamiliar target, onboarding a new team member, screening candidates in a scouting or BD role, or preparing a target for de novo binder design. Also serves as a demonstration of domain-informed MCP server design for Anthropic's life sciences ecosystem.

Design Philosophy

1. Tool descriptions are prompt engineering. Claude reads the docstrings and Field descriptions to decide when/how to use each tool. Every description should encode scientific reasoning, not just data labels. This is where domain expertise matters most.

2. Composition over coverage. 8 tools that compose into a real workflow > 30 isolated database wrappers. The tools are ordered to reflect how expert medicinal chemists actually evaluate targets.

3. Pre-compute interpretations. Each tool returns raw data AND an interpretation field with scientific context. This helps Claude reason without dumping raw JSON into its context.

4. Fail gracefully with useful messages. Biological databases are flaky. Missing data is informative (no ChEMBL data = untargeted = opportunity). Errors should tell the scientist what the absence means, not just "404."

5. Filter noise. Crystallization artifacts (glycerol, PEG, sulfate) pollute binding site analysis. The artifact filter list is scientifically important.

Architecture

pocketscout-mcp/
├── CLAUDE.md              ← You are here
├── pyproject.toml         
├── README.md              ← Design rationale (important for portfolio)
├── src/pocketscout_mcp/
│   ├── server.py          ← MCP server: 8 tools + 2 prompts
│   ├── models.py          ← Pydantic models (field descriptions = MCP schema)
│   └── clients/           ← Async API wrappers
│       ├── base.py        ← Shared HTTP client with retries
│       ├── uniprot.py     ← Protein annotation, function, family
│       ├── pdb.py         ← Structures, ligands, binding sites
│       ├── alphafold.py   ← Structure confidence (pLDDT)
│       ├── chembl.py      ← Bioactivity, competitive landscape
│       └── pubmed.py      ← Literature search
├── tests/
│   └── test_clients.py    
└── examples/
    └── egfr_assessment.md ← Example walkthrough