# LabClaw: 240 biomedical skills written as markdown files an agent reads at runtime

> A skill library rather than a software package, LabClaw ships SKILL.md files that teach an OpenClaw agent when to call a bioinformatics or drug discovery tool and what output to expect.

**wu-yc/LabClaw** — LabClaw – Operating Layer for LabOS (Stanford-Princeton AI Co-Scientists)

- Repository: https://github.com/wu-yc/LabClaw
- Website: https://labclaw-ai.github.io/
- Stars: 1,055 · Forks: 161
- Language: Unknown
- License: not declared
- Published: 2026-10-07 · Updated: 2026-10-07 · Language: en
- Canonical page: https://hysenlabs.com/projects/wu-yc-labclaw

## A skill library, deliberately not a package

LabClaw is 240 `SKILL.md` files in one `skills/` directory. There is no build step, no dependency manifest at the root, and no compiled artifact. The repository tree is short enough to read in full: `README.md`, a Chinese-language `README.zh-CN.md`, a logo image, two demo GIFs, and `skills/`.

The README states the design position directly. LabClaw is a skill library, not a monolithic software package, and you can install the whole collection or copy only the folders that match your workflow. Each skill teaches an OpenClaw-compatible agent when to use a tool, how to call it, and what kind of output to produce. That last clause is the one that separates this from a prompt collection. A prompt tells a model what to do; a skill also tells it what a correct result looks like, which is the part that usually gets left implicit and then guessed wrong.

The project describes itself as an operating layer for LabOS, the Stanford and Princeton AI co-scientist work, with the stated goal of powering dry-lab reasoning, protocol composition and agentic workflows that close the loop with XR or wet-lab execution. The homepage is labclaw-ai.github.io, and the repository description carries the same framing.

One licensing detail deserves a second look. The README badge says MIT and links to a `LICENSE` file, but no `LICENSE` file appears in the top-level tree listing, and the licence field GitHub reports for this repository is empty. If the licensing of a set of instructions you plan to redistribute matters to your group, confirm the licence text exists before shipping a copy.

## Installing one skill from the agent chat window

The quick start is a single message sent to OpenClaw. The README calls it a three-second install:

```bash
install https://github.com/wu-yc/LabClaw
```

There is no clone-and-copy step in the documentation because the runtime does the fetching. The related-projects table is explicit that openclaw/openclaw is the main runtime that loads workspace skills and provides the skills platform, the onboarding flow and the agent workspace model LabClaw is designed to fit into. So the agent is the delivery mechanism, not a convenience wrapper around a manual copy.

That also means the manual path is the fallback rather than the documented path. Nothing in the README explains how to register a skill directory with OpenClaw by hand, which is the question anyone running a self-hosted or pinned runtime will ask first.

The repository includes donation addresses for BNB and Solana, immediately followed by a line stating no crypto affiliation and that all memecoins are scams. The disclaimer is there to pre-empt impersonation rather than to solicit anything, and it is worth knowing the addresses are visible if you ever see them attached to something claiming to be this project.

## What the seven domain folders actually contain

The README publishes a domain table, and the numbers are the fastest way to understand the shape of the collection.

Biology and life sciences is the largest folder at 86 skills, covering bioinformatics, single-cell analysis, genomics, proteomics, multi-omics and database access. General and data science is next at 54, spanning statistics, machine learning, data management, scientific writing and quality control. Pharmacy and drug discovery holds 36 skills across cheminformatics, molecular machine learning, docking, target research and pharmacology. Literature and search has 33, aimed at academic search, biomedical databases, multi-source discovery, patents, grants and citations. Medical and clinical has 22, covering clinical trials, precision medicine, oncology, infectious disease and medical imaging. Vision and XR is the smallest of the substantive folders at 5, for hand tracking, 3D pose estimation, segmentation and egocentric vision. Visualization holds 4, aimed at matplotlib, seaborn and plotly.

Those add to 240, which matches the badge. One inconsistency is worth flagging: the collapsible catalog further down the README labels the biology folder as 66 skills rather than 86. Either the catalog is stale or the badge is. Nobody documents which, so the count for that folder should be verified by listing the directory rather than trusted from the page.

The README then gives representative workflows, which are more informative than the raw counts because they show how skills compose. Single-cell and spatial omics maps to `anndata`, `scanpy` and `tooluniverse-spatial-transcriptomics`. Drug discovery maps to `rdkit`, `diffdock` and `tooluniverse-drug-repurposing`. Clinical work maps to `clinical`, `tooluniverse-precision-oncology` and `clinicaltrials-database`.

## Where LabClaw sits against ToolUniverse and Biomni

The README names three related projects and explains the division of labour for each, which is more useful than a bare list of links.

openclaw/openclaw is the runtime, and the distinction matters: LabClaw contains no runtime code, so the two are complementary rather than competing. mims-harvard/ToolUniverse is described as a large AI-scientist tool ecosystem, and LabClaw ships many skills named with a `tooluniverse-` prefix across omics, drug discovery, clinical workflows and literature research. In other words the tool implementations live there and the instructions for calling them live here. snap-stanford/Biomni is a complementary biomedical AI agent project, and LabClaw already includes a `biomni` skill.

That third relationship is the one to think about. A skill that teaches an agent to drive another agent system is a different proposition from a skill that calls a Python library, because the failure mode moves up a layer: when the advice is wrong, the inner system may still produce a plausible-looking result. Skills built on top of ToolUniverse have the same property.

The practical consequence is that LabClaw is best read as a mapping layer over tools you already have access to. It does not vendor rdkit, scanpy or anndata, and nothing in the tree suggests it installs them. A skill telling an agent to use `diffdock` is only useful if a working docking setup is reachable from the agent's environment.

## Reading a skill file before you trust it

The unit of this project is a markdown file, which has an upside and a downside. The upside is that markdown is diffable, reviewable and editable without any tooling, so you can change what an agent is told to do by editing a sentence in a text file. The downside is that the same property means a skill can be wrong in plain language, with nothing to fail a test.

Given that, the way to evaluate the collection is to read the skills you plan to use rather than the aggregate. Pick one folder, open two or three `SKILL.md` files, and check for three things. First, whether the tool paths and function signatures in the examples match the version of the library you have installed, since an API renamed between versions produces an agent that fails confidently. Second, whether the expected-output description is specific enough to be checkable, because a skill that says the output should be reasonable gives an agent nothing to verify against. Third, whether the skill names its failure cases, which is a decent proxy for whether it was written by someone who has run the workflow.

The domain table's grouping makes this practical. Nobody installs 240 files to evaluate them, but copying one folder, at most 86 files, is cheap to review and cheap to delete. That is the intended usage pattern according to the README, and it is also the right way to find out whether the quality is uniform before committing a whole lab to it.

## Maintenance state and the count mismatch to check first

The last push to this repository was on 2026-03-19. The repository is not archived and has 4 open issues, but the gap between that date and now is long enough that the collection should be treated as a fixed snapshot rather than a moving target. Nothing in the README documents a release cadence, and the repository has no GitHub releases at all, so there is no version marker to pin against other than a commit hash.

For a library whose content is instructions rather than code, staleness has a specific failure mode. A library dependency gets a patch release and your build breaks loudly. A skill file gets edited upstream and your agent quietly changes behaviour on the next install, with no error and no diff in your own repository unless you pinned a commit. If reproducibility matters to your work, copy the folders you use into your own workspace rather than referencing the upstream repository, which also solves the licensing question above.

The second thing to verify is arithmetic. The domain table sums to 240, matching the headline badge, while the catalog section reports 66 for biology against the table's 86. A twenty-skill gap in one folder is large enough to suggest either an unreflected addition or an unremoved section. Count the directories under `skills/bio/` yourself before assuming any particular skill exists, and do the same for the folders you actually plan to use.

## Conclusion

LabClaw is worth installing if you run an OpenClaw agent for biomedical work and have been writing tool instructions from scratch one skill at a time, since the value is that 240 of those instructions already exist and can be cherry-picked by folder rather than adopted wholesale. It is not worth adopting as a general-purpose prompt bundle, and it will not help if your agent runtime is not OpenClaw-compatible, because the skill format is the product and there is nothing else here to install. Verify two things before you rely on it: that the counts in the badges match what you actually get, since the README shows both 86 and 66 for the biology folder, and that the upstream tools the skills call are installed on your side, since a skill teaches an agent how to invoke something that has to exist. The last push to the repository was on 2026-03-19, so pin a commit rather than tracking the default branch if a skill silently changes under you.

## FAQ

### What is LabClaw and what does it actually contain?

LabClaw is a skill library rather than a software package. It contains 240 SKILL.md files under a skills/ directory, grouped into bio, vision, pharma, med, general, literature and visualization folders, and each file teaches an OpenClaw-compatible agent when to use a tool, how to call it and what output to expect.

### How do you install LabClaw?

The README documents a single message sent to OpenClaw: install https://github.com/wu-yc/LabClaw. There is no clone-and-copy step because the OpenClaw runtime loads workspace skills, and the repository does not document a manual registration path.

### Does LabClaw need ToolUniverse or Biomni installed separately?

The skills call tools rather than shipping them. The README lists mims-harvard/ToolUniverse as the AI-scientist tool ecosystem that LabClaw's tooluniverse-prefixed skills target, and notes that LabClaw already includes a biomni skill for snap-stanford/Biomni. Neither is vendored into this repository.

### Can I install only part of the LabClaw collection?

Yes, and the README recommends it. It states you can install the full collection or copy only the skill folders that match your research workflows, with biology the largest at 86 skills and visualization the smallest at 4.

### Is LabClaw still being updated?

The repository is not archived and its last push was on 2026-03-19, with 4 open issues. It has no GitHub releases, so there is no version tag to track and pinning a commit is the way to keep a skill set stable.

## Sources

- [Issues](https://github.com/wu-yc/LabClaw/issues)
- [Project website](https://labclaw-ai.github.io/)
- [README](https://github.com/wu-yc/LabClaw/blob/main/README.md)
- [wu-yc/LabClaw on GitHub](https://github.com/wu-yc/LabClaw)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/wu-yc-labclaw