Frequently asked questions

Don’t see what you’re looking for here? Check the Discourse or use search_knowledge in the MCP server.

Why doesn’t the MCP server run my model?

By design — this is a knowledge-only server. It returns documentation, code templates, and guidance. To actually train a model, install scvi-tools and run the code that the server provides.

How do I keep the knowledge base up to date?

GitHub Actions jobs handle this automatically and open PRs when diffs are found:

  • refresh_knowledge.yaml — re-scrapes GitHub issues + Discourse

  • sync_tutorials.yaml — syncs tutorial notebooks from scvi-tools

  • sync_model_knowledge.yaml — updates model docs from the scvi-tools CHANGELOG

  • sync_huggingface_hub.yaml — refreshes the Hugging Face Hub model registry quarterly

You can also trigger them manually via workflow_dispatch or run the scripts locally:

python scripts/scrape_external.py
python scripts/convert_notebooks.py --src /path/to/scvi-tutorials --dst src/scvi_tools_mcp/knowledge/tutorials
python scripts/extract_api_docs.py
python scripts/scrape_huggingface_hub.py

Which models are covered?

Run get_model_overview or list_tutorials in the MCP server. The knowledge base currently covers scVI, scANVI, TotalVI, MultiVI, PeakVI, LinearSCVI, MrVI, Stereoscope, CellAssign, SOLO, GIMVI, veloVI, ContrastiveVI, SCBASSET, SysVI, DiagVI, and more.

For pretrained model repos on the official scvi-tools Hugging Face organization, use list_hub_models, get_hub_model, or suggest_hub_models.

What is the uvx command?

uvx runs a package from PyPI without installing it permanently — provided by uv. It is the recommended way to run MCP servers because it always fetches the latest published version.

Can I use this server from R?

Yes — any MCP-compatible client works. Configure your client to run uvx scvi-tools-mcp and use the tools from whatever language your client supports. The server itself is language-agnostic.