tiangzlab/omicsclaw

OmicsClaw turns local multi-omics tools into AI-callable skills. The LLM plans and operates; Python, R, and CLI tools process your data in a local or remote runtime — raw matrices never leave your machine. One agent loop serves the terminal, the desktop app, and chat platforms.

📢 What’s New

• 🤝 Consensus runtime — multi-method consensus is now a declarative workflow runtime. Fan out N spatial-clustering or single-cell methods, then merge them with verified typed operators or an exploratory LLM synthesis. Triggered by the consensus-domains and sc-consensus-clustering skills.
• 🧠 Autonomous Analysis Path — an Analysis Router can parameterize an exact skill from your data, or run a generated-code analysis with approval-gated workspace writes and bounded LLM repair.
• ⚡ Prompt-prefix caching — automatic provider cache hits across turns to cut latency and token spend.
• 🖥️ Desktop upgrades — a live to-do task list with planning guidance, an interactive ask_user choice tool, and LLM-generated session titles.

🖥️ App Workspace

One workspace for chat, datasets, skills, execution, memory, and analysis outputs.

📥 Download the OmicsClaw Desktop App  ·  All releases  ·  SHA256SUMS

The Releases tab hosts the prebuilt desktop installers — the same oc desktop-server the CLI ships, wrapped in a chat-ready Electron UI. Pick the asset for your platform:

Verify each download against SHA256SUMS.txt published alongside the installers. The desktop client and the CLI talk to the same backend — analyses, memory, and remote runtimes stay portable across both.

💡 Why OmicsClaw?

✨ Capabilities

⚡ Quick Start

git clone https://github.com/TianGzlab/OmicsClaw.git
cd OmicsClaw
bash 0_setup_env.sh
conda activate OmicsClaw
oc list
oc run spatial-preprocess --demo

Configure chat and runtime settings:

oc onboard
oc interactive

If oc is not on PATH, use python omicsclaw.py .

🧭 Interfaces

Pick the entry point that fits your workflow — they all reach the same backend.

Remote mode uses 127.0.0.1, SSH tunneling, and OMICSCLAW_REMOTE_AUTH_TOKEN. See remote execution and the legacy remote guide.

Channels are Owner-only: Feishu additionally requires FEISHU_ALLOWED_SENDERS and FEISHU_BOT_OPEN_ID (the identity that proves a group message mentioned this bot). Everything not listed above — other adapters, outbound media — fails closed.

📦 Installation

📖 Details: installation guide, quickstart. Dependencies live in pyproject.toml, environment.yml, and 0_setup_env.sh.

🚀 npm install & Desktop pairing

One npm install -g omicsclaw gives you the CLI and a self-contained CPython runtime — no conda, no venv, no system Python. That same runtime is the interpreter the Desktop App can be pointed at, so a single install serves both the terminal and the App.

Status — the wrapper and its four platform runtimes are built by npm-release.yml; publishing is a manual, reviewer-gated dispatch that has not run yet, so npm install -g omicsclaw still 404s on the registry. Until it lands, install the backend through the conda or pip paths above.

npm install -g omicsclaw   # CLI + the one runtime matching your platform
omicsclaw --version        # `oc` is the short alias for the same binary
oc list                    # 95 skills, by domain

Node.js 18+ is the only prerequisite. The wrapper carries no runtime: it declares one @omicsclaw/runtime- per host in optionalDependencies, and npm’s os / cpu filtering lands exactly one on disk — the pattern esbuild and biome use. The postinstall hook records that interpreter in ~/.omicsclaw/runtime.json and renames any pip-installed omicsclaw / oc shim to -legacy, so the npm command wins PATH without deleting the old one.

The runtime carries the agent and the desktop server, not the scientific stack (scanpy, torch, R, bioconda CLIs — roughly 1.5 GiB). Skills that need those tell you what to install into the same interpreter; for the full supported stack, use the Linux conda path.

Pairing with OmicsClaw-App

The desktop installer contains no Python, and never downloads, creates, repairs, or auto-selects an interpreter. You pick one explicitly; the App commits it only after a preflight, a provisional launch, and a strict /health check, and restores the previous runtime if any of that fails.

Print the exact interpreter path when the App asks for one:

# npm runtime
python -c "import json, os; print(json.load(open(os.path.expanduser('~/.omicsclaw/runtime.json')))['pythonPath'])"
# conda env
conda run -n OmicsClaw python -c "import sys; print(sys.executable)"

The backend binds 127.0.0.1:8765 (OMICSCLAW_APP_HOST / OMICSCLAW_APP_PORT); remote profiles authenticate with OMICSCLAW_REMOTE_AUTH_TOKEN. Configure the LLM provider in the App’s setup wizard or in the backend’s .env. Chat-triggered runs are written to /output, which is what the App dashboard lists.

📖 Distribution internals — wrapper layout, platform packages, and the ~/.omicsclaw/runtime.json contract — are documented in npm/AGENTS.md and npm/omicsclaw/README.md.

🧬 Domains

oc list and skills/catalog.json currently agree on 95 registered skills across 8 domains.

Run oc list for the current CLI catalog.

🧠 Memory

Graph-backed memory at omicsclaw/memory/ carries your sessions, datasets, analyses, preferences, and insights across runs — chat history and lineage come back when you reopen any surface. Each surface stays isolated so state never leaks across users or workspaces.

A reserved __shared__ pool (core agent identity, knowledge handbook guards, glossary) is the one thing every surface reads back automatically. Full vocabulary and architecture in docs/CONTEXT.md.

📚 Documentation

Hosted docs site: ****

❓ FAQ

⚠️ Safety

OmicsClaw is a research and educational tool for multi-omics analysis. It is not a medical device and does not provide clinical diagnoses. Consult a domain expert before making decisions based on these results.

See data privacy and rules/disclaimer.

👥 Community

Maintainers: Luyi Tian (Principal Investigator), Weige Zhou (Lead Developer), Liying Chen (Developer), and Pengfei Yin (Developer).

🐛 Issues · 💬 Discussions · 📖 Docs

🙏 Acknowledgments

Architecture, skill design, and local-first philosophy are inspired by ClawBio, an early bioinformatics-native AI agent skill library. Memory and session-continuity patterns are inspired by Nocturne Memory.

🛠️ Contributing

• New skills: see CONTRIBUTING.md and the v2 scaffold under templates/skill/.
• Repository / agent work: see AGENTS.md — covers contract tests, provider contracts, skill runner, and architecture references.

📜 License

Apache-2.0. See LICENSE.

📝 Citation

@software{omicsclaw2026,
  title = {OmicsClaw: A Memory-Enabled AI Agent for Multi-Omics Analysis},
  author = {Zhou, Weige and Chen, Liying and Yin, Pengfei and Tian, Luyi},
  year = {2026},
  url = {https://github.com/TianGzlab/OmicsClaw}
}

⬆ Back to top