dcc-mcp-slicer

A typed 3D Slicer adapter for DCC MCP. Create bounded mathematical volumes, material phantoms and sphere meshes, inspect and edit them, then export NRRD/STL or round-trip a native MRB scene bundle through MCP. Configure native cameras and orthogonal slices, then render bounded PNG previews.

Current candidate: trusted local synthetic workflows with exact dcc-mcp-core 0.20.41 / dcc-mcp-server 0.20.41 dependencies. Native Slicer, graphical camera/render and native MCP cancellation acceptance must be rerun for this candidate. The earlier Linux native evidence used Slicer 5.10.0 / Python 3.12.10 / Core and server 0.20.39; it does not qualify the upgraded runtime. See the runtime upgrade and rollback record. Synthetic data only; no patient data, medical interpretation, clinical claims, DICOM import or arbitrary scene-file loading.

Quick start

Install this checkout with Slicer's bundled PythonSlicer, not system Python:

/path/to/Slicer/bin/PythonSlicer -m pip install /path/to/dcc-mcp-slicer

Set up a dedicated, trusted artifact directory, then run in a fresh Slicer Python console:

import os
from dcc_mcp_slicer import start_server

os.environ["DCC_MCP_SLICER_WORKSPACE"] = "/absolute/existing/synthetic-artifacts"
server = start_server()  # OS-assigned loopback port; gateway disabled by default
print(server.mcp_url)

Keep server alive. Call server.stop() from the same application/main thread to shut down. A stopped server/dispatcher is terminal; call start_server() again to create a fresh instance. See installation and removal, architecture, and reproducible validation.

Related MCP server: rupa3d

MCP workflow

Initialize a normal MCP Streamable HTTP client at the printed URL:

1. search_skills({"query":"slicer synthetic"})

2. load_skill({"skill_name":"slicer-scene"})

3. Call the returned canonical tool slug, for example: slicer_scene__create_phantom({"size":64})

4. Use returned node IDs with the edit/export tools

5. Poll Core's jobs_get_status with the returned job ID for asynchronous calls

Prefix each local tool name with slicer_scene__. Core may advertise shortened aliases or paginate/cap tool listings; the load_skill response contains canonical registered slugs. This package does not implement its own MCP wire protocol.

Boundaries

• MRML, VTK and Qt calls execute on the Slicer application thread

• Startup rejects unsupported Slicer versions and missing workspaces. Synthetic node creation/import is capped at 100 nodes

• Node readback includes a session-local revision. Pass expected_revision to sphere edit/export to reject stale requests; reacquire node IDs after reload

• Output paths are plain basenames in the configured workspace; traversal, symlinks and overwrite are rejected. Native writes use private staging and exclusive atomic publication; failed/cancelled exports do not publish partial files

• Only adapter-marked synthetic nodes can be edited/exported; scene saves reject non-synthetic storable data, including hidden storable nodes

• MRB imports have size, archive-path, member-type and external-reference checks

• The workspace must be operator-controlled and trusted. A hash receipt detects accidental changes; it does not authenticate third-party content

• Native save/load/export calls are asynchronous but cannot be interrupted after entering Slicer. Poll the existing job; never replay a timed-out mutation

• Trusted loopback clients can also access Core administration/discovery surfaces. This adapter is not a sandbox for hostile clients or third-party skills

• PNG rendering requires a working native graphics context and active graphical layout. This captures only Slicer view buffers, never the desktop or other apps

• Windows/macOS and Slicer versions outside the 5.10 family are not supported by this profile. Earlier native acceptance used 5.10.0 with runtime 0.20.39; the 0.20.41 candidate still requires the same native acceptance procedure

Development

python -m pip install -e '.[dev]'
python -m ruff check src tests scripts
python -m ruff format --check src tests scripts
python -m pytest
python -m build --outdir /tmp/dcc-mcp-slicer-dist

CI runs source, closed input/typed output schemas, the official MCP SDK, real loopback HTTP and fake-timer tests. Real Slicer acceptance is a separate executable smoke; it is not silently replaced by mocks. No package release or automated publishing workflow is enabled.