omicsclaw

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

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

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

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

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.

Documentation

The docs site is written in Chinese.

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