google-deepmind/science-skills: an agent skill bundle for genomics, structural biology and literature search

What problem Science Skills solves, and who it is aimed at

An agent asked to look up a gene, pull a protein structure and check the literature will happily invent endpoints. Science Skills addresses that by shipping pre-written instructions for each scientific task, so the agent follows a documented path instead of improvising one. The README describes the bundle as "a collection of agent skills for scientific research tasks, spanning genomics, structural biology, cheminformatics, literature search, and more", and says each skill provides "structured instructions, scripts, and resources that extend an AI agent's capabilities for specialized scientific tasks".

The audience is narrow and specific. It is not a teaching resource, despite what the phrase "science skills" suggests to most search engines. It is for people running an AI agent who want that agent to reach AlphaGenome, AFDB, UniProt and the other integrated sources with better grounding and, per the project description, higher token efficiency. If you are writing a Python data pipeline by hand, this bundle gives you nothing to import. If you are operating an agent that reads skill directories, it gives you the domain knowledge you would otherwise have to write yourself.

The SKILL.md directory layout and how an agent consumes it

The repository is a flat container of skill directories under skills/, plus CONTRIBUTING.md, LICENSE, README.md, SKILL_LICENSES.md and plugin.json at the top level. Each skill directory follows the same three-part shape, as the README lays out: SKILL.md, a scripts/ directory, and an optional references/ directory.

SKILL.md carries YAML frontmatter and detailed markdown instructions. That frontmatter is what an agent runtime reads to decide whether the skill is relevant to a request, which is why the format matters more than the prose inside. The scripts/ directory holds helper scripts and utilities the skill invokes, and references/ holds extra documentation. The Python primary language of the repository lives in those scripts rather than in a single installable package.

This is a convention-driven design, not a framework. Nothing in the README describes a loader, a registry or a plugin API. The agent runtime is responsible for finding skill directories, parsing the frontmatter and executing the scripts. That means the bundle's behaviour depends heavily on which agent you run it under, and the README only documents one runtime in detail: Google Antigravity.

Installing Science Skills with npx skills add

The README gives a single install command for the bundle, using npx. Run it from a terminal where Node and npx are available:

npx skills add google-deepmind/science-skills/

After the command completes you should have the skill directories on disk in the location your agent runtime reads. The README does not spell out that destination, so check your runtime's documentation before assuming where they landed.

For Google Antigravity there is a second path. New users download Antigravity and tick the Science box at the 'Build with Google' step, which the README says installs "the curated collection of our Science Skills". Existing users update Antigravity, then open Settings -> Customizations -> Build with Google Plugins, click 'Customize' at the bottom of the page, and download the Science plugin.

Dependencies are handled by uv. The README states that the first time you trigger a Science Skill, the agent asks for approval, installs uv, and then proceeds with the task. It recommends restarting Antigravity after that first installation. Some skills need credentials. AlphaGenome and OpenAlex require an API key; ClinVar works without one but gets higher rate limits with it. The README says the agent should prompt you for the key and show you where to write it, but if you prefer to do it yourself it gives this command:

echo "ALPHAGENOME_API_KEY=your_actual_api_key" >> ~/.env

A first real use looks like this: install the bundle, restart Antigravity if this was the first uv bootstrap, then ask the agent a genomics question that maps to a skill such as AlphaGenome or ClinVar. Expect an approval prompt for uv on the first trigger and, for keyed skills, a prompt to supply the key before any result comes back.

Custom skills belong outside the plugin directory

The README is unusually direct about one failure mode: do not edit files inside the Antigravity Science plugin installation directory. Changes there are overwritten whenever the plugin updates. The supported alternative is to place custom or modified skills in your personal skills directory:

~/.gemini/config/skills/

This is a real constraint rather than a style preference. Because the plugin ships as a unit and updates as a unit, any local patch to a shipped SKILL.md has a limited lifetime. If your workflow depends on a modified version of a shipped skill, you need to maintain a copy under ~/.gemini/config/skills/ and keep it in sync yourself. The README does not describe a merge or override mechanism, so how a personal skill of the same name interacts with the shipped one is undocumented. Test that before you rely on it.

Where Science Skills is the wrong tool

The bundle has no standalone library surface. There is no documented import path, no published package name for the Python code, and no CLI of its own. The scripts/ directories are invoked by the agent following SKILL.md, not by you. If your requirement is a deterministic, testable pipeline that runs in CI without an agent in the loop, this is the wrong shape of project, and you would be better served calling the underlying databases directly.

Runtime coupling is the second limitation. The documented install and customization paths assume Antigravity, down to the ~/.gemini/config/skills/ directory name. The npx skills add route is runtime-agnostic in principle, but the README does not state which runtimes read the resulting layout, so support elsewhere is unverified.

Credentials are the third. AlphaGenome and OpenAlex skills require an API key to function at all, so an agent that cannot prompt for or read a key environment variable will stall on those skills. The README notes that the agent should guide you through the key setup, but that guidance is itself agent behaviour, not a guarantee.

Finally, the disclaimer is explicit that this "is not an official Google product". Treat support expectations accordingly.

Alternatives and the difference in approach

The closest alternative is writing your own tool wrappers against the same upstream sources. If you register AlphaGenome, UniProt and OpenAlex as tools in your agent framework directly, you control the request schema, the retry policy and the token budget, and you can unit test each wrapper. Science Skills trades that control for ready-made domain instructions: the SKILL.md files encode which fields matter for a given scientific task, which is knowledge that is tedious to reconstruct and easy to get subtly wrong. The cost of the trade is that you inherit the project's structure and its update cycle.

A second alternative is a general-purpose agent tool server that exposes HTTP APIs without domain guidance. That gives broader coverage of arbitrary endpoints but no scientific grounding, which is the specific thing this bundle claims to add. The distinction is instruction content, not connectivity: both approaches let an agent make a call, but only one tells it what a good call looks like for genomics or structural biology.

The README points to a technical report on the Science Skills for readers who want the reasoning behind the design rather than the mechanics.

Licence, maintenance and upgrade cost

The software is Apache-2.0, and the README states that all other materials are under CC-BY 4.0. That split matters because the skills are mostly prose and data references, which fall under the Creative Commons side rather than the patent-granting software licence.

More important for adoption is SKILL_LICENSES.md, which the README describes as 'Skill Licences and Terms of Use'. It says certain third party data sources referenced within individual Skill files have their own applicable licences or terms of use, and that you are responsible for ensuring your use of individual Skill files complies with them. This is not a formality: the bundle integrates AlphaGenome, AFDB, UniProt and 30+ other sources, and each may carry different terms. Read SKILL_LICENSES.md for the specific skills you plan to use before you build a workflow on them. Nothing here is legal advice, and the file is the authoritative statement.

On maintenance, the last push to main was on 2026-09-15, two days before this writing, and v1.2.0 was released on 2026-09-08. The repository is not archived. The upgrade cost is concentrated in two places. Plugin updates overwrite anything you edited inside the plugin directory, which is why custom work belongs in ~/.gemini/config/skills/. And because the bundle has no lockfile story in the README, pinning a known-good set of skill files means copying them out and tracking upstream changes yourself.

Editorial conclusion

Adopt Science Skills if you already work inside Google Antigravity or an agent that reads SKILL.md directories, and you want AlphaGenome, UniProt, OpenAlex and similar sources reachable without writing your own tool wrappers. Do not adopt it if you need a standalone Python library you can import, or if you cannot supply the API keys that AlphaGenome and OpenAlex require. Before committing, verify three things: whether your agent runtime reads ~/.gemini/config/skills/, which of your target skills need a key, and what SKILL_LICENSES.md says about the third party data sources behind the skills you plan to use.

Frequently asked questions

Official sources

1. google-deepmind/science-skills on GitHub
2. License: Apache-2.0
3. Project website
4. README
5. Releases