DeepMind's agent toolkit for wet-lab drudgery

What it does

Science Skills is a collection of modular instruction sets for AI agents doing specialized scientific work—genomics, structural biology, cheminformatics, literature search. Each skill is a directory with a SKILL.md file (YAML frontmatter plus detailed markdown instructions), helper scripts, and optional references. The agent reads the instructions, installs dependencies via uv, and queries the relevant database. Think of it as a cookbook where every recipe includes not just ingredients but the exact knife technique.

The interesting bit

The grounding angle: instead of hoping an LLM hallucinates the right UniProt query syntax, you give it a pre-vetted skill with correct endpoints, rate-limit handling, and API key prompts. DeepMind claims this improves “token efficiency”—less back-and-forth correction, fewer expensive retries. The skills are installable via npx skills add or bundled into Google’s “Antigravity” agent platform.

Key highlights

• Covers 30+ databases and tools including AlphaGenome, AFDB (AlphaFold DB), UniProt, ClinVar, OpenAlex
• Each skill is self-contained: instructions, scripts, and references in one directory
• Dependency management through uv; first run triggers install with user approval
• Some skills require API keys (AlphaGenome, OpenAlex); others work without but rate-limit you (ClinVar)
• Apache 2.0 for code, CC-BY for other materials; third-party data sources have their own licenses (see SKILL_LICENSES.md)
• Published technical report available (linked in README)

Caveats

• README is vague on what “better grounding” or “higher token efficiency” actually means numerically—no benchmarks cited
• “Not an official Google product” disclaimer present despite DeepMind branding
• Antigravity integration is the primary path; standalone usage outside that ecosystem is underdocumented

Verdict

Worth a look if you’re building scientific agents and tired of debugging why your LLM keeps querying the wrong ClinVar endpoint. Skip it if you want a general-purpose RAG framework—this is tightly scoped to molecular biology and chemistry workflows, and the real value assumes you’re already in the Antigravity orbit.