OmicsClaw AI Agent

<a id="top"></a>

<div align="center"> <a href="https://github.com/TianGzlab/OmicsClaw"> <img src="docs/images/OmicsClaw_banner.jpeg" alt="OmicsClaw banner" width="100%"/> </a> <h3>Local-first AI research partner for multi-omics analysis</h3> <p> <b>English</b> · <a href="README_zh-CN.md">简体中文</a> · <a href="#quick-start">Quick start</a> · <a href="#desktop-app">Desktop App</a> · <a href="#domains">Domains</a> · <a href="https://TianGzlab.github.io/OmicsClaw/">Docs site</a> </p>

</div>

OmicsClaw is an AI agent for multi-omics analysis. You describe the analysis you want; the model picks a skill, reads its instructions and runs its Python, R or command-line tools in your own environment, on your machine or on a server. It ships 88 skills in seven domains, and the same agent answers in the terminal, in the desktop app and on chat platforms.

What's new

• Across seven domains, 83 skills expose function libraries and replayable examples; five remain CLI-only, including the demo-only XCMS stub, and Python/R modules support IO tracking and review before acceptance (0074, 0075, 0076).
• A real-model routing eval (tests/evals/live/, run by hand) measures whether the model picks the right skill for 26 seed requests, the CI unit-test job now also runs the launch shell, attachments, top-level and desktop HTTP tests, and the four consensus skills whose scripts could not start have since been deleted (0068, 0070).
• The agent framework has been rebuilt around three entry points, oc cli, oc desktop and oc channel, and the old oc interactive, oc tui, oc onboard and oc run commands are gone (rebuild status).
• Desktop v0.1.8 is available for macOS, Windows and Linux, with Desktop v3, SSH session recovery and an optional Independent review button below each completed reply.
• The agent tells you which of a skill's packages your environment lacks, and with OMICSCLAW_SKILL_ENV=install it can install them after you approve, into a separate overlay that leaves the base environment untouched (0061).

Earlier entries are in CHANGELOG.md.

Quick start

git clone https://github.com/TianGzlab/OmicsClaw.git
cd OmicsClaw
bash 0_setup_env.sh        # creates the OmicsClaw conda environment
conda activate OmicsClaw
oc cli --configure         # choose a model provider and enter its API key
oc cli                     # chat in the terminal

oc cli --configure saves its answers to a .env file. It asks for one of 13 providers (DeepSeek, OpenAI, Anthropic, Gemini, Qwen on DashScope, Ollama, any OpenAI-compatible endpoint and others), the API key, the model, an optional base URL, and the workspace directory the agent may read and write. The agent loads skills from <workspace>/skills, so if the workspace is not the checkout, also set OMICSCLAW_SKILLS_DIR=<checkout>/skills in .env. oc cli --prompt "..." answers one question and exits. If oc is not on PATH, use python omicsclaw.py cli.

Skills are ordinary scripts, so you can also run one without the agent:

make list    # every skill, by domain
python skills/spatial/spatial-preprocess/spatial_preprocess.py --demo --output /tmp/omicsclaw_demo

Ways to use it

Where	Start it with	What you get
Terminal	oc cli	Chat with the agent. /resume reopens an earlier conversation, /skills lists skills, /help lists the other commands
Desktop App	the App starts oc desktop	Chat, approvals and file viewing in a desktop window (below)
Remote server	oc desktop on the server	The App reaches it through its own SSH tunnel, so the data stays on the server (remote mode)
Chat platforms	oc channel --channels telegram	Owner-only text chat on Telegram, Feishu, Slack, Discord, DingTalk, QQ or Email. oc channel --list shows them. Set OMICSCLAW_APPROVAL_TIMEOUT_S in .env first, or it refuses to start
Skill scripts	python <skill directory>/<script>.py	One analysis without the agent. Each skill's SKILL.md documents its options

Conversations and the agent's long-term notes are stored in <workspace>/.omicsclaw/memory.db. MCP servers listed in <workspace>/.mcp.json are connected at start-up, and their tools become available to the agent.

The agent asks you before it runs a shell command, writes or edits a file, goes to the web, or calls an MCP tool. In the terminal you can allow a call once, allow that tool for the rest of the conversation, or always allow that exact call; /auto stops the questions about ordinary work. Dangerous commands and changes to .omicsclaw/ or .env are still asked about. The dangerous-command check matches known patterns, and a command written to avoid them gets through, so if you let the agent work unattended, run its commands in a container without network access by setting OMICSCLAW_SANDBOX=docker and OMICSCLAW_SANDBOX_IMAGE (see .env.example).

Desktop App

<p align="center"> <img src="docs/images/omicsclaw-app-overview.png" alt="The OmicsClaw desktop app, with a project's conversations in the sidebar and the chat box in the middle" width="94%"/> </p>

Download Desktop v0.1.8 for your platform below. Check each download against the SHA256SUMS.txt published with it.

Platform	Installer
macOS, Apple Silicon	OmicsClaw-0.1.8-arm64.dmg
macOS, Intel	OmicsClaw-0.1.8-x64.dmg
Windows, x64 or ARM64	OmicsClaw.Setup.0.1.8-x64.exe, OmicsClaw.Setup.0.1.8-arm64.exe
Linux, x64	AppImage, deb, rpm
Linux, ARM64	AppImage

The installer contains no Python. Remote mode connects to a server through SSH or a backend URL and needs no Python on the client. Local mode asks for an interpreter with OmicsClaw installed, normally the conda environment from the quick start:

conda run -n OmicsClaw python -c "import sys; print(sys.executable)"

In Local mode, the App starts oc desktop for the project you open, one backend per project. The model comes from the backend's .env, which the App's Providers page can edit; a change takes effect when the backend restarts. Troubleshooting is in the App guide, and the HTTP contract in docs/core-features/surfaces.md §8.

Independent module review is off by default. After a reply finishes, click Independent review below it to review that reply's analysis and return the findings in the same conversation. The click applies only to that review; it does not enable reviews for later analyses. Use App v0.1.8 or later, update the backend to include 1540fca7, and restart oc desktop to use this control.

Installation

Method	Command	Covers
conda (recommended)	bash 0_setup_env.sh	Python, R and the command-line tools the skills need, plus the Desktop backend
pip	pip install -e .	Terminal chat only. Most skills need packages that the conda environment provides, and the Desktop backend's FastAPI and uvicorn come from conda

Chat platforms need their own SDKs: pip install -e ".[channels]" for Telegram and Feishu, pip install slack-sdk aiohttp for Slack, pip install discord.py for Discord, pip install httpx websockets for DingTalk and pip install qq-botpy for QQ; Email uses only the standard library. The dependency lists are environment.yml and pyproject.toml.

Domains

88 skills in seven domains. make list prints the current index.

Domain	Skills	Examples	Guide
Spatial transcriptomics	17	QC, domains, annotation, deconvolution, CNV, trajectory	spatial
Single-cell omics	30	QC, clustering, annotation, doublets, velocity, GRN	singlecell
Genomics	10	QC, alignment, variants, CNV, assembly, epigenomics	genomics
Proteomics	8	DIA/DDA, PTM, networks, biomarkers	proteomics
Metabolomics	8	Peaks, normalization, annotation, pathways	metabolomics
Bulk RNA-seq	14	DE, enrichment, co-expression, deconvolution, survival, cosinor rhythms	bulkrna
Literature	1	PDF, DOI, PubMed and GEO parsing, dataset handoff

Documentation

The docs site is written in Chinese.

Topic	Where
Quickstart	introduction/quickstart
Desktop App	ecosystem/omicsclaw-app
Remote mode	engineering/remote-execution
Domain guides	spatial · singlecell · genomics · proteomics · metabolomics · bulkrna
Safety and data privacy	data privacy · rules and disclaimer
Writing a skill	CONTRIBUTING.md · templates/skill/
Framework design	docs/FRAMEWORK-REBUILD.md
Changelog	CHANGELOG.md

Safety and data

Skills read and process your data in your own runtime, local or remote; the model receives the conversation and what the tools return. OmicsClaw is a research tool. It is not a medical device and gives no clinical diagnoses, so have a domain expert check results before acting on them. On a server, keep oc desktop bound to localhost and reach it through SSH; a non-loopback bind requires OMICSCLAW_REMOTE_AUTH_TOKEN. See data privacy and rules and disclaimer.

Community

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

Issues · Discussions · Docs

<table> <tr> <td align="center" width="30%"> <img src="docs/images/IMG_3729.JPG" alt="OmicsClaw WeChat group QR code" width="180"/> <br/> WeChat group </td> <td valign="middle" width="70%"> Scan the code to join our WeChat group for analysis tips and help with problems. </td> </tr> </table> <a href="https://github.com/TianGzlab/OmicsClaw/graphs/contributors"> <img src="https://contrib.rocks/image?repo=TianGzlab/OmicsClaw" alt="OmicsClaw contributors"/> </a>

Acknowledgments

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

Contributing

To add a skill, see CONTRIBUTING.md and the scaffold under templates/skill/. For work on the framework itself, start with AGENTS.md.

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