Unofficial MCP server for Google's Neuroglancer volumetric viewer. Lets Claude (or any MCP client) drive a Neuroglancer session: load datasets, navigate, show/hide segments, search by published metadata, style for figures, share snapshot URLs.

This project is not affiliated with Google. Neuroglancer is developed by the Connectomics at Google team.

Status: pre-release, not yet on PyPI. Install from source (below). PyPI publication is gated on stabilizing the tool surface and a courtesy heads-up to the Neuroglancer maintainers.

Neuroglancer is the standard viewer for petascale connectomics and EM segmentation data (FlyEM hemibrain, MICrONS, FlyWire, H01, CellMap). It has a rich Python API but no agentic interface. This server exposes that API as MCP tools so an LLM can compose viewer operations in response to natural-language requests — e.g., "load the hemibrain and show me three Kenyon cells" becomes a sequence of add_segmentation_layer + show_segments calls.

This project uses uv for environment and dependency management. The lockfile (uv.lock) is committed; reproduce the dev environment with:

git clone https://github.com/janelia-cellmap/neuroglancer-mcp.git
cd neuroglancer-mcp
uv sync                       # creates .venv, installs runtime + dev deps
uv run pytest                 # run the test suite
uv run pytest -m network      # also run live-network tests (hits public GCS buckets)
uv run neuroglancer-mcp       # start the MCP server on stdio

There is no PyPI release yet — pip install neuroglancer-mcp will not work.

Until PyPI publication, MCP clients launch the server via uv run against the cloned project. Substitute /abs/path/to/neuroglancer-mcp for your checkout path.

claude_desktop_config.json (Claude Desktop):

{
  "mcpServers": {
    "neuroglancer": {
      "command": "uv",
      "args": ["--directory", "/abs/path/to/neuroglancer-mcp", "run", "neuroglancer-mcp"]
    }
  }
}

.mcp.json at the repo root (Claude Code, project-scoped — already shipped in this repo):

{
  "mcpServers": {
    "neuroglancer": {
      "command": "uv",
      "args": ["run", "neuroglancer-mcp"]
    }
  }
}

Reload your MCP client. Ask: "Load the FlyEM hemibrain dataset and show me a Kenyon cell."

The first tool call returns a viewer URL. Open it once in your browser; all subsequent calls update that viewer live.

State & introspection

• get_state — current viewer state as JSON
• get_url — live viewer URL (only reachable while this MCP server runs)
• share_url — snapshot URL on neuroglancer-demo.appspot.com that encodes the current state for sharing
• save_share_url — write the snapshot URL to a file (HTML or plain text) for when it's too long to inline in chat
• load_state — load state from JSON or share URL
• list_layers — names, types, visibility

Navigation

• navigate_to — move to (x, y, z) in nm or voxels
• set_zoom — set cross-section scale
• get_position — current center position
• set_orientation — set the 3D view quaternion
• center_on_layer — re-center on a layer's volume midpoint (precomputed + OME-NGFF zarr); useful if you used center=False earlier or navigated away
• center_on_segment — move the viewer to a single segment's approximate centroid (reads the mesh manifest; supports sharded multilod_draco and legacy unsharded meshes)

Layers

• add_image_layer — explicit image layer; auto-centers on the volume by default (center=False to opt out)
• add_segmentation_layer — explicit segmentation layer; same auto-center behavior
• add_layer — smart add: infers image vs segmentation from the source's dtype (float* → image, uint64 → segmentation, others → image with low confidence)
• add_layers — bulk add (one transaction, one auto-center); each item's type is optional and auto-detected when omitted
• remove_layer, set_layer_visibility
• set_layers_visibility — bulk-toggle visibility by {name: bool} map
• show_only_layers — show the named layers, hide everything else (the "isolate this view" shortcut)

Segments

• show_segments — replace visible segments in a layer
• add_segments — add to current selection
• hide_segments — remove from selection
• clear_segments — hide all

segment_properties (search and filter by published metadata)

• get_segment_properties_schema — what properties this dataset publishes
• get_segment_property — one value for one segment
• search_segments_by_label — substring or regex over the label property
• filter_segments — combine label glob, tags, and numeric ranges
• top_n_segments_by_property — e.g. N largest neurons by precomputed size
• show_segments_by_property — filter and display in one round trip

Styling — colors and opacity

• set_segment_colors — per-segment hex overrides ("make neuron 1234 red")
• set_layer_color — default color for a segmentation layer
• set_layer_opacity — 0–1 opacity (image opacity, or 3D object_alpha for segs)
• set_mesh_silhouette — silhouette-edge intensity for figure-quality meshes

Styling — view arrangement

• arrange_view — switch among Neuroglancer's nine native layouts (xy, 4panel, 3d, …) or semantic aliases (quad, 3d_only, top_xy)
• set_background_color — cross-section and/or 3D background (e.g. white for figures)

Styling — image shaders

• set_image_shader — raw GLSL on an image layer
• set_image_preset — grayscale, red/green/blue, fire, viridis, threshold
• set_contrast — convenience window: maps [min, max] to [0, 1] grayscale

Styling — segmentation rendering

• set_segmentation_rendering — voxel, mesh, or both
• set_mesh_resolution — mesh LOD threshold (smaller = finer)

Local data

• serve_local_directory — explicitly expose a filesystem path over HTTP so Neuroglancer's browser can read it
• list_served_directories, stop_serving_directory — manage the spawned servers
• Auto-served: passing a bare filesystem path (e.g. /groups/cellmap/.../foo.zarr) to any add_*_layer tool spawns the server automatically, detects the format from extension or contents, and rewrites the source URL

The viewer binds to 0.0.0.0 by default, so get_url returns a hostname-based URL that other machines on the same LAN can open (typical workflow: MCP runs on a workstation, you open the viewer on your laptop). To restrict to the local machine, set NEUROGLANCER_MCP_BIND_ADDRESS=127.0.0.1 before launching; you can also pass a specific interface IP for tighter firewalling.

share_url is independent of bind address — it always emits a neuroglancer-demo.appspot.com snapshot.

Note: the viewer URL contains a random per-session token, so it isn't guessable, but it has no auth on top of that. Don't use the default bind on untrusted networks.

• Every tool returns the resulting state fragment so the agent can verify its action.
• Coordinate units are explicit on every navigation call (nm vs voxels) — the model should never guess.
• The server holds one viewer instance per process. Multi-viewer support is deferred.

BSD 3-Clause. See LICENSE.

• neuroglancer — the upstream viewer
• napari-mcp — analogous MCP server for napari
• cloud-volume — Python client for Neuroglancer Precomputed data