Every ClawBio skill is a folder: a SKILL.md contract, the script that executes it, demo data and tests. Claude Code installs the library as a plugin. Cursor, VS Code, Codex and Zed now read the same folders directly. So the MCP server we shipped in July is retiring, 41 pull requests of hardening ship with it, and on 25 September we are at the Broad to build agents with you.
The skills have not changed shape. What changed is who can read them. The Agent Skills format ClawBio adopted at the start is now native in the editors people actually use, so the supported routes are these.
The plugin ships the skills themselves. The manifests now track the package version, enforced by a test, after sitting at 0.3.0 while PyPI shipped 0.6.1.
/plugin marketplace add ClawBio/ClawBio /plugin install clawbio
All four read ~/.agents/skills/. Copy or symlink the folders you need. Each is self-contained.
git clone https://github.com/ClawBio/ClawBio
cp -r ClawBio/skills/pharmgx-reporter \
~/.agents/skills/
The CLI and the API return the same structured result and reproducibility bundle the MCP tools returned.
pip install --upgrade clawbio clawbio run pharmgx --demo
9am to 4pm ET at 300 Binney Street. One day, real genomic questions, agents that do useful scientific work, then show the room what they found. Hosted with Mo Tanha. Seats are limited.
Contributors, skill authors and the people running the hackathons are in one workspace. Ask a question, share a skill, or find a team for Cambridge.
We built it in 0.6.0 for a room in Kendall Square that ran Cursor, Zed and VS Code and had no way to use ClawBio. That gap has since closed from the other side: Cursor 2.4, VS Code 1.108, Codex and Zed 1.4.2 all load Agent Skills from ~/.agents/skills/ without any server.
A remote-call shim in front of folders the editor already reads adds a process, a dependency pin (mcp<2) and a second code path, for no capability the skills lack. One contract to keep correct is better than two. The contract is the SKILL.md and the script beside it.
bioqc-mcp, bgpt-mcp, just-prs-mcp) are unaffected. This is about ClawBio's own server only.SHA-256 helpers live in clawbio.common.checksums. The scaffold now generates a working bundle for every new skill, and equity-scorer, nutrigx and gwas-prs write one. Every analysis can be replayed without the agent that produced it.
deepspot-m: virtual spatial transcriptomics from H&E tiles. just-prs-mcp: evidence-aware polygenic scores from VCF or WGS through a pinned local just-prs server.
pharmgx-reporter had three GRCh37 positions in a GRCh38 table. nutrigx had the rs4988235 lactose polarity inverted. clinical-variant-reporter now fails closed when a VEP batch fails. gwas-prs demo panels no longer shadow real PGS Catalog scores. rnaseq-de keeps gene IDs across the nf-core handoff and checks that LFC shrinkage matches the requested contrast.
The tests of any skill a PR touches now run. Skills-only PRs from new contributors are approved automatically. The scientific audit gates on a committed benchmark baseline, so a regression fails the build instead of a number in a report.
Skills and Galaxy tools from skills/catalog.json; tests counted from def test_ definitions in the repository; stars and contributors from the GitHub API on 3 September 2026.
Six external contributors landed code between 0.6.1 and 0.7.0. Reproducibility bundles, the new skills and most of the clinical fixes came from them.