clawsight

name	clawsight
description	Operate WSInsight whole-slide pathology AI via its Docker MCP server

ClawSight Skill

ClawSight gives you full control over WSInsight — an end-to-end whole-slide image (WSI) pathology analysis toolkit — by connecting to a WSInsight MCP server running inside a Docker container.

Pipeline overview

WSI file  →  wsinsight_patch  →  wsinsight_infer  →  wsinsight_ncomp  →  wsinsight_export
              (tissue seg +        (GPU model            (Delaunay graph      (GeoJSON /
               HDF5 patches)        inference)            neighborhood)        OME-CSV)

Or run everything in one call with wsinsight_run.

Quick-start

1. wsinsight_start_docker({ "data_dir": "/path/to/slides", "gpu_ids": "0" })
2. # wait ~5 seconds
3. wsinsight_connect({})
4. wsinsight_list_tools({})           ← discover exact parameter names
5. wsinsight_run({ "arguments": { "wsi_dir": "slides",
                                  "results_dir": "results",
                                  "model": "breast-tumor-resnet34.tcga-brca" } })
6. wsinsight_job_status({ "job_id": "<id>" })  ← poll until "done"
7. wsinsight_stop_docker({})          ← clean up when finished

Rules

• File paths inside arguments must be relative to /workspace (= the data_dir you passed to wsinsight_start_docker).
• Argument names are the WSInsight CLI parameter names in snake_case (e.g. wsi_dir, results_dir, batch_size, num_workers, region_inference_dir, export_geojson). Always call wsinsight_list_tools first — the MCP server exposes the canonical schema and the plugin does not hard-code it.
• Async tools (long-running on the server) return job_id immediately: run, patch, infer, ncomp, plus the experimental hplot, ecomp, tcomp, cme. Poll wsinsight_job_status until status is "done" or "error".
• Sync tools (export, reg, and the experimental hplot-finalize, cme-profile) block until completion and return output directly.
• If you get a connection error, call wsinsight_connect to re-establish the session (the server may have restarted or the container may have cycled).
• Experimental tools (hplot, hplot-finalize, ecomp, tcomp, cme, cme-profile) appear in wsinsight_list_tools only when the container was started with "experimental": true (which sets WSINSIGHT_EXPERIMENTAL=1 and launches the server with --experimental).

Tool reference

Tool	Category	Blocks?
wsinsight_server_info	Connection	sync
wsinsight_connect	Connection	sync
wsinsight_start_docker	Docker	sync
wsinsight_stop_docker	Docker	sync
wsinsight_list_tools	Discovery	sync
wsinsight_run	Pipeline	async → job_id
wsinsight_patch	Pipeline	async → job_id
wsinsight_infer	Pipeline	async → job_id
wsinsight_ncomp	Pipeline	async → job_id
wsinsight_export	Pipeline	sync
wsinsight_reg	Pipeline	sync
wsinsight_job_status	Job mgmt	sync
wsinsight_job_logs	Job mgmt	sync
wsinsight_cancel_job	Job mgmt	sync
wsinsight_list_jobs	Job mgmt	sync

Available models

Pass one of the names below as the model argument. The MCP server resolves them from the bundled WSInsight zoo registry inside the container; no network access is required.

Cell-level (object-based) models — each cell becomes one row in model-outputs-csv/<slide>.csv:

• CellViT-256-x20, CellViT-256-x40, CellViT-256-x40-AMP
• CellViT-SAM-H-x20, CellViT-SAM-H-x40, CellViT-SAM-H-x40-AMP
• CellViT-Virchow-x40-AMP
• 10xGenomics-BRCA-CellViT-SAM-H-x40, 10xGenomics-CRC-CellViT-SAM-H-x40
• hovernet_fast_pannuke
• hne_cell_classification

Region / patch-level models — each patch becomes one row (useful as region_inference_dir for reg or for region_prob_* columns):

• breast-tumor-resnet34.tcga-brca
• lung-tumor-resnet34.tcga-luad
• pancreas-tumor-preactresnet34.tcga-paad
• prostate-tumor-resnet34.tcga-prad
• pancancer-lymphocytes-inceptionv4.tcga
• lymphnodes-tiatoolbox-resnet50.patchcamelyon
• colorectal-tiatoolbox-resnet50.kather100k
• colorectal-resnet34.penn

Selection rules:

• The x20 / x40 suffix on CellViT models must match the slide magnification (TCGA diagnostic SVS slides are typically 40x).
• --model is mutually exclusive with the trio (--config + --model-path) and --zoo-model-dir (folder with config.json + torchscript_model.pt). For ad hoc weights, pass them via the config/model_path or zoo_model_dir arguments instead of model.

Output data formats

Everything lands under the results_dir you passed (relative to /workspace):

<results_dir>/
  masks/<slide>.jpg                  Tissue segmentation thumbnails
  patches/<slide>.h5                 Patch coords (and optional images)
  model-outputs-csv/<slide>.csv      Per-cell (or per-patch) inference table
  ncomp-outputs-csv/<slide>.csv      Per-cell neighborhood composition
  graphs/<slide>.h5                  Cached Delaunay graph
  export-csv/<slide>.csv             Merged per-cell table (model + ncomp)
  export-geojson/<slide>.geojson     QuPath-compatible GeoJSON
  export-omecsv/<slide>.ome.csv.gz   QuPath / OMERO+ compatible OME-CSV
  patch_metadata_<ts>.json           Patch-stage configuration
  infer_metadata_<ts>.json           Inference-stage configuration

model-outputs-csv/<slide>.csv (per-cell or per-patch)

Columns:

• minx, miny, width, height — bounding box in level-0 pixels.
• prob_<class> — one float column per model class. The class names come from the model's bundled config.json (e.g. prob_tumor, prob_lymphocyte).
• (object-based models only) center_x, center_y — cell centre in level-0 pixels.
• (when region_inference_dir is supplied) region_minx, region_miny, region_width, region_height, region_prob_<class> — enclosing region patch and its class probabilities. Argmax of region_prob_* gives a per-cell region label (e.g. tumor vs non-tumor).

ncomp-outputs-csv/<slide>.csv (per-cell composition)

• center_x, center_y — cell centre.
• cell_type — argmax over prob_* from the model output.
• neighborhood_size — number of k-hop neighbours (excluding self).
• neighborhood_<type>_count and neighborhood_<type>_prop — per-class counts and proportions across the k-hop neighbourhood. _prop is NaN when neighborhood_size == 0.

export-csv/<slide>.csv

Left-join of model-outputs-csv/ with ncomp-outputs-csv/ on (center_x, center_y). Same columns as the two sources combined.

patches/<slide>.h5 (HDF5)

• /coords — (N, 2) int32, top-left [x, y] of each patch at level 0. Attributes: patch_size, patch_level, patch_spacing_um_px, optional tile_dim.
• /slide.attrs — slide_path, slide_mpp, slide_width, slide_height.
• /images — (N, patch_size, patch_size, 3) uint8, only when the run used cache_image_patches=True.
• /polygons/{coords, offsets} — ragged polygon vertices (when polygons were supplied).

graphs/<slide>.h5 (HDF5, produced by ncomp)

• cell_centers — (N, 2) int32.
• simplices — (M, 3) int32 Delaunay triangle vertex indices.
• edges_source, edges_target, edges_length — unpruned undirected edges (length in pixels).
• file.attrs — num_cells, mpp, centers_hash (SHA-256 of cell_centers.tobytes() for cache invalidation).

Edges are stored unpruned; pruning to ncomp_max_neighbor_distance happens at read time.

export-geojson/<slide>.geojson

Standard GeoJSON FeatureCollection. Each feature:

{
  "type": "Feature",
  "id": "<uuid4>",
  "geometry": {"type": "Polygon", "coordinates": [[[x1,y1], ...]]},
  "properties": {
    "isLocked": true,
    "objectType": "detection",       // or "tile" / "annotation"
    "classification": {"name": "prob_<winner>", "color": [R,G,B]},
    "measurements": {"prob_tumor": 0.92, "neighborhood_tumor_prop": 0.7, ...}
  }
}

measurements includes every numeric column except the geometry columns (minx, miny, width, height, center_x, center_y).

export-omecsv/<slide>.ome.csv.gz

Gzip-compressed CSV with columns:

• object — row index.
• secondary_object — same as object.
• polygon — WKT polygon string.
• objectType — "detection" / "tile" / "annotation" (chosen via object_type argument to wsinsight_export).
• classification — argmax class name with the prob_ prefix stripped.
• All numeric non-geometry columns. NaN is written as the literal string "NaN".

Reading results from agent code

import pandas as pd
df = pd.read_csv("results/model-outputs-csv/SLIDE.csv")
print(df.columns.tolist())   # incl. prob_<class> for every model class