Scientific Skills Desktop (Bun MCPB)

A Claude Desktop extension (.mcpb) that delivers the claude-scientific-skills corpus via a tiny librarian tool surface. Built as a self-contained Bun binary (no Node/Python required for users). Offline fat bundle by default; optional thin/auto-update mode.

Features

• 128+ curated scientific skills, JIT-loaded to avoid prompt bloat
• Tools: search_skills, read_skill, list_skills
• Domains: bioinformatics, cheminformatics, proteomics, clinical, materials, physics, ML
• Data sources: PubMed, ChEMBL, COSMIC, UniProt, ClinicalTrials.gov (26+)
• Libraries: RDKit, Scanpy, DeepChem, DiffDock, scvi-tools, Qiskit (54+)
• Integrations: Benchling, OMERO, Protocols.io (15+)
• Offline fat bundle; optional thin/auto-update via CORPUS_URL

Install (Desktop)

1. Download scientific-skills-fat.mcpb (fat/offline) from Releases.
2. Claude Desktop → Settings → Extensions → Install Extension → select the .mcpb.

Usage

// Search for a skill
{ "tool": "search_skills", "arguments": { "query": "biopython" } }

// Read a skill (full doc)
{ "tool": "read_skill", "arguments": { "id": "bioinformatics/biopython/SKILL.md" } }

// Read a section with truncation
{ "tool": "read_skill", "arguments": { "id": "bioinformatics/biopython/SKILL.md", "section": "Installation", "max_chars": 4000 } }

Thin/auto-update mode (optional)

• Set CORPUS_URL (env) to a tar.gz corpus; server will download to ${__dirname}/.cache on first run.
• CORPUS_PATH and CACHE_DIR are set via manifest env to the bundle root by default.

Build

bun build src/index.ts --compile --outfile bin/skills-mcp
./scripts/pack-fat.sh   # fat bundle with corpus
./scripts/pack-thin.sh  # thin bundle (no corpus)

Credits

• Corpus and original MCP concept: K-Dense-AI/claude-scientific-skills (MIT)
• Desktop packaging and Bun binary: Mat Silverstein

License

MIT