TotalSegmentator: a skill for choosing segmentation tasks
The official skill guides an agent through MCP capability discovery, task, device and speed selection, then records the generated masks and run settings.
The complete kit in one ZIP
One ZIP: skills, instructions, installer and a public demo runner.
When to use it
Select exact classes and a compatible task
Prepare segmentation of a public CT sample
Record settings, masks and limitations of the statistics
What you’ll get
An explicit operation choice and a result record; masks are produced only by a real run of the separately installed MCP server.
For example
Prepare segmentation of a public CT sample: use MCP to check available classes and tasks and choose the device and speed. Before running, state the settings and expected downloads. Keep masks and statistics separate from medical conclusions.
Before you start
Our guide to the official TotalSegmentator skill. The article’s earlier CLI run does not validate this MCP server or skill.
Read the pinned SKILL.md and separate MCP documentation. Register skills/totalsegmentator from the reviewed revision using your agent’s supported mechanism.
Prepare an isolated TotalSegmentator environment with the mcp extra and a compatible device. Check Python/PyTorch, weight storage and selected-task terms before installing.
Connect a local stdio server. The documented HTTP mode is unauthenticated; do not expose it on the internet.
Use MCP to discover classes and settings, then state the device, speed, compatible ROI and statistics choices. Initial inference can download weights.
Use a permitted public sample and preserve settings, masks and raw output. Check alignment and incomplete structures; model outputs are not clinical findings.
Original, requirements and limits
Original skill and software: the TotalSegmentator team, University Hospital Basel. Editorial guide: kossolapov.com.
- Agent
- An Agent Skills-compatible agent with an MCP client
- Tools
- Python, uv or pipx, PyTorch and a separately installed TotalSegmentator[mcp] environment
- A compatible CPU/GPU/MPS device, RAM, disk space and access to model weights
- A permitted public NIfTI sample or supported DICOM input and an image viewer for mask review
- Access
- Some tasks require a separate TotalSegmentator license; inspect the selected task first
- Possible costs
- Licensed tasks and rented compute can incur costs; the total task is listed as Apache-2.0 in the reviewed README
- Limits
- This pinned skill and MCP server were source-reviewed, not executed. The earlier CLI 2.18.0 sample run is separate evidence.
- MCP HTTP mode has no authentication in the reviewed documentation; prefer local stdio, or restrict HTTP to loopback.
- The first inference can download weights. Paths refer to the server machine.
- Upstream documents anonymous usage reporting and its send_usage_stats setting; inspect data flows before use.
- An empty pathology mask does not exclude disease; incomplete structures can distort statistics. Outputs are estimates, not clinical findings.