OpenClaw Medical Skills: 869 medical agent skills for OpenClaw and NanoClaw

What OpenClaw Medical Skills adds to a general-purpose agent

A Claude-based assistant with no domain modules answers medical questions from its training data. It cannot query PubMed, pull a trial record from ClinicalTrials.gov, or run a variant annotation pipeline, because it has no interface to those systems. OpenClaw Medical Skills is a collection of 869 SKILL.md files that supply those interfaces. The README frames the difference in a table: without skills, generic responses about medicine; with them, real PubMed, ClinicalTrials.gov and FDA queries, RNA-seq and scRNA-seq pipelines, ChEMBL and DrugBank lookups, SOAP notes and discharge summaries, VCF annotation with ACMG classification, and regulatory guidance for FDA, CE mark, IEC 62304 and ISO 14971.

The audience is narrow and worth naming. This is for people who already run OpenClaw or NanoClaw and want to extend it, not for clinicians looking for a finished product. The repository has no application, no UI and no server. It is a directory of markdown skill definitions plus a plugin manifest, and the value depends entirely on the host agent reading them. If you are evaluating medical AI tools generally, this is the wrong entry point; if you have an agent and want it to stop hallucinating database results, it is the right one.

How a SKILL.md file turns into a database query

Each skill is a self-contained module. According to the README, a SKILL.md file teaches the agent specialized domain knowledge and workflows, connects it to real databases, APIs and computational tools, and produces structured clinical or scientific output. The mechanism is prompt-level, not code-level: the host agent discovers the skill directory, reads the markdown, and follows the described workflow, calling whatever tools the skill names.

That design has a consequence the README does not discuss. Because the skill is instructions rather than a compiled adapter, the reliability of a PubMed query depends on the agent following the instructions correctly on that run. There is no schema validation between the skill and the external service. The collection is organized into categories, with counts given in the README: 10 general and core skills, 119 medical and clinical, 43 scientific databases, 239 bioinformatics skills under the gptomics label, and 59 omics and computational biology skills covering single-cell and spatial work. The README states the collection aggregates skills from more than 12 open-source skill repositories, which means the quality and maintenance of individual skills varies with their upstream source. The top-level layout matches this: a skills/ directory, a scripts/ directory, openclaw.plugin.json, and the two READMEs.

Installing OpenClaw Medical Skills with sparse checkout

The README gives four installation methods. The recommended one avoids downloading bundled data files by using a sparse checkout that fetches only the skills directory. Run this from a directory where you want the clone to live:

git clone --depth=1 --no-checkout https://github.com/FreedomIntelligence/OpenClaw-Medical-Skills.git
cd OpenClaw-Medical-Skills
git sparse-checkout init --cone
git sparse-checkout set skills
git checkout main

After the checkout completes you have a skills/ directory without the large databases and datasets that some skills bundle. Copy it into your workspace skills path, which the README lists as the high-priority location, or into the global path shared by all agents:

cp -r skills/* <your-workspace>/skills/
# or, for all agents:
cp -r skills/* ~/.openclaw/skills/

The README states skills are picked up automatically on the next session and that no restart is needed. To confirm, ask the agent what medical and clinical skills it has available; the README says it should list the installed skills with their capabilities.

If you want only part of the collection, the README shows a loop over an explicit skill list. This example installs a clinical and drug-discovery stack:

SKILLS=(
  "clinical-reports"
  "tooluniverse-drug-research"
  "tooluniverse-pharmacovigilance"
  "clinicaltrials-database"
  "biomedical-search"
  "tooluniverse-drug-drug-interaction"
)

for skill in "${SKILLS[@]}"; do
  cp -r OpenClaw-Medical-Skills/skills/$skill ~/.openclaw/skills/
done

There is also a configuration route. Adding the cloned repository to ~/.openclaw/openclaw.json mounts the whole collection without copying files:

{
  "plugins": {
    "local": ["/path/to/OpenClaw-Medical-Skills"]
  }
}

For NanoClaw the target is different. The README says NanoClaw loads skills into agent containers at startup from container/skills/, so you clone, copy into that path, and rebuild:

git clone https://github.com/FreedomIntelligence/OpenClaw-Medical-Skills.git
cp -r OpenClaw-Medical-Skills/skills/* /path/to/nanoclaw/container/skills/
cd /path/to/nanoclaw
./container/build.sh

One caveat on the clone command itself: the README uses the full clone for NanoClaw and for the case where you want bundled data, which means Git LFS should be installed first if those data files matter to you.

Where the collection breaks down

The most concrete limitation is licensing. The README carries an MIT badge, but the repository's top-level entries do not include a LICENSE file, and the licence is listed as unknown. An MIT badge in a README is not the licence text. For a collection that aggregates skills from more than 12 upstream repositories, the licence of each skill may differ from the badge, and nothing in the README addresses per-skill provenance. If you plan to use these skills in a commercial or regulated setting, that gap is the first thing to resolve, and it is not something the README can resolve for you.

The second limitation is the install surface. Some skills bundle large data files, which is why the recommended method uses sparse checkout. That protects your bandwidth but also means a sparse install may be missing data a specific skill expects. The README does not document what happens when a skill runs without its bundled data, and it does not document rollback or uninstall beyond deleting the copied directories.

The third is scope mismatch. The collection is broad by design: 869 skills across clinical documentation, bioinformatics, drug discovery, regulatory compliance and general tooling. If you work in one narrow area, most of what you install is noise in the agent's context. The README's Method 4 exists precisely because bulk installation is not always what you want. And because skills are prompt instructions rather than validated adapters, a skill that queries a live database can fail silently if the upstream service changes its interface; the README documents no error handling for that case.

OpenClaw Medical Skills compared with a single-purpose medical agent

The obvious alternative is a purpose-built medical AI tool that ships its own retrieval layer and UI, rather than a skill library you attach to a general agent. The difference in approach is where the integration lives. A dedicated tool owns the pipeline end to end: it controls how a PubMed query is formed, how results are parsed, and how the answer is presented. OpenClaw Medical Skills pushes that work into markdown instructions the host agent interprets at runtime.

That trade-off cuts both ways. The skill library is composable: you can install clinical-reports and tooluniverse-drug-drug-interaction side by side, add biomedical-search, and get a combined workflow no single-purpose tool offers. It is also inspectable, since a SKILL.md file is readable text you can edit. What you give up is determinism. A dedicated tool returns the same structured output for the same input; a skill-driven agent may not, because the model decides how to follow the instructions on each run. If your use case needs reproducible output for audit, the skill approach is the weaker choice. If it needs breadth and you are willing to review outputs, it is the stronger one.

Maintenance, upgrades and what the licence badge does not cover

The last push to the repository was on 2026-07-21, roughly two months before this writing, and the repository is not archived. There are no retrieved releases, so upgrades are not versioned in a way you can pin. The README describes openclaw plugins update as the command to update all installed skills, which implies updates flow through the plugin mechanism rather than through tagged releases of this collection. If you installed by copying directories, re-running the copy is your upgrade path, and nothing in the README describes how to detect that an upstream skill changed.

On licensing, the README shows an MIT badge and links to a LICENSE path, but the repository listing does not surface a LICENSE file and the licence is recorded as unknown. MIT is permissive and would allow commercial use with attribution, but that is a statement about the badge, not about the files. Because the collection aggregates from more than 12 upstream repositories, each skill could carry its own terms. Check the LICENSE file and the individual skill directories before relying on the badge. This is not legal advice; it is a description of what the repository does and does not state.

Editorial conclusion

Adopt it if you already run OpenClaw or NanoClaw and want your agent to reach PubMed, ClinicalTrials.gov, FDA and bioinformatics pipelines without writing each integration yourself; start with the sparse-checkout method and copy only the handful of skill directories your domain needs. Do not adopt it if you expect a standalone medical application, if you cannot accept an MIT badge with no LICENSE file content to read, or if your work is regulated in a way that requires provenance for every tool in the chain. Verify three things before you commit: the actual text of the LICENSE file, whether the skills you selected bundle external API keys or data downloads, and whether your OpenClaw version resolves the skills/ directory as the README describes.

Frequently asked questions

Official sources

1. FreedomIntelligence/OpenClaw-Medical-Skills on GitHub
2. Issues
3. README