Set up CellTypist as a remote tool

Sharing this model for a user, not developing it? Follow the tooluniverse-share-computer guide; no ToolUniverse source checkout is needed. The model runs in its own Python environment, with its dependencies:

~/.local/bin/uv venv --python 3.12 ~/.tooluniverse/envs/celltypist
curl -fsSL -o ~/.tooluniverse/envs/celltypist-tooluniverse.txt https://connect.aiscientist.tools/install-requirements.txt
curl -fsSL -o ~/.tooluniverse/envs/celltypist-requirements.txt https://raw.githubusercontent.com/mims-harvard/ToolUniverse/main/src/tooluniverse/remote/celltypist/requirements.txt
~/.local/bin/uv pip install --python ~/.tooluniverse/envs/celltypist/bin/python -r ~/.tooluniverse/envs/celltypist-tooluniverse.txt -r ~/.tooluniverse/envs/celltypist-requirements.txt
~/.local/bin/tu remote check celltypist --python ~/.tooluniverse/envs/celltypist/bin/python   # lists anything still missing
~/.local/bin/tu serve --allow celltypist --python ~/.tooluniverse/envs/celltypist/bin/python --share --detach --name "<computer name>" --service https://tooluniverse-backend.onrender.com

If the check reports a missing environment variable (a data folder, a model cache, a provider key), set it in the same shell before tu serve; "Obtain credentials, data, and model weights" below says what each one is. The model reads only the variables it declares.

The rest of this guide is for developers working from the ToolUniverse source, and for diagnosing problems with this model.

Validation status (2026-08-16): working CPU deployment. The live MCP call annotated 80 cells from the official CellTypist sample and completed majority voting. A converted data-only model produced the same 80 labels and probabilities as the digest-pinned upstream model (maximum probability delta 0.0). This is runtime equivalence evidence, not an annotation-accuracy benchmark. Authenticated private Platform import and owner testing passed on 2026-08-16; public publication and independent-caller authorization/isolation remain untested.