CBCTScope MCP Server
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
CBCTScope
A complete, local-first CBCT and 2D radiograph viewer, built by a board-certified oral and maxillofacial radiologist.
It can also be driven by an AI agent over MCP, under one rule: the agent moves the camera and never produces findings; the human reads.
Website (in your browser, nothing installed) · Run it (three steps) · Desktop app (macOS, Windows) · Open your own scan · Reading guides · User manual · Agent control (MCP) · Cite it
![]()
Above: a real CBCT in MPR. Axial, sagittal and coronal sections cross at the skull base and the sphenoid sinus in the bone window, beside a bone 3D render of the facial skeleton.
Clinicians and students: read the CBCT exports already on your computer in eight purpose-built modes, free, with no account, no upload, and nothing leaving the machine. It starts on a built-in phantom, so the whole viewer can be learned with zero patient data.
Researchers: one instrument that reads CBCT volumes and 2D radiographs under identical display settings, for multi-reader studies, with NIfTI, STL, and PNG slice-stack export and a citable DOI for every release.
Agent builders: a built-in MCP server with fifteen navigation-only verbs, a one-click bundle for MCP hosts such as Claude Desktop, and an Agent Skill. The agent moves the camera; the human reads.
CBCTScope opens the CBCT exports already on your computer and reads them in eight purpose-built modes: MPR with 3D rendering, multi-slice grid, curved panoramic with nerve tracing, TMJ, freehand reslice, virtual cephalogram, region growing with airway analysis, and two-volume stitching. It reads 2D radiographs (panoramic, intraoral, CR and DX) in the same instrument, under the same display settings. It is free, open source, and local-first: no account, no cloud, no upload, ever. No agent is required; CBCTScope is a complete standalone viewer.
Through a built-in MCP server, an agent can open scans, switch reading modes, set the window, move through slices, and take snapshots, and nothing more: it never produces findings. The agent moves the camera. The human reads. That split was the design from the first commit. To our knowledge, CBCTScope is the first purpose-built CBCT reading viewer with native AI-agent control; the neighboring projects, and how they differ, are under Why this exists.
Research use only. Not a medical device. CBCTScope visualizes and navigates volumes. It never produces findings, measurements-as-conclusions, or diagnoses, in the UI or over MCP. Do not use it for clinical decision-making.
The interface at a glance
MPR + 3D | Multi-slice grid | Curved panoramic | TMJ condyles |
|
|
|
|
The four stills above are a real CBCT: a 16×16cm, 200 µm Planmeca ProMax volume.
Click any thumbnail for full resolution. Each reading mode has its own guide, written the way a reader actually moves through a volume: docs/reading-modes. The full user manual, from installation to the keyboard reference, lives in docs/manual.
Related MCP server: MedVision MCP
Run it
Three steps, no technical background needed. CBCTScope is not an installed app with an icon: it is a small local program you start once per reading session, and your browser is the screen. Everything is undone by deleting the folder.
System requirements
Node.js 22 or later: the free runtime that runs the viewer, the way a browser runs a web page.
A current desktop browser with WebGL2 (Chrome, Edge, Firefox, or Safari). The 3D render and all reformats are computed on this machine, so a recent computer with a working graphics driver is assumed; large volumes read more comfortably with more memory.
macOS, Windows, or Linux.
Install Node.js 22 or later from nodejs.org: download the LTS installer, run it, accept the defaults.
Get CBCTScope. Click the green Code button at the top of this page, choose Download ZIP, and unzip it somewhere you can find again (Desktop is fine). If you use git,
git cloneworks as usual.Start it. In the unzipped folder, double-click Start CBCTScope.command on macOS or Start CBCTScope.bat on Windows. If the system shows a security prompt on this first open, the one-time answer for each system is under "If something goes wrong" below. The first start downloads the viewer's components once (about 130 MB downloaded, 600 MB on disk: ten to fifteen seconds on the author's Mac and connection in September 2026, longer on a slow one); then your browser opens the viewer by itself. A built-in geometric test phantom loads immediately: simple geometric shapes with an arch of tooth-like cylinders, a test object for trying the controls, not anatomy. The full viewer works with zero patient data, and the status line says so until you open a scan of your own. Help in the header opens the reading guide for the mode you are in.
To stop the viewer, close the starter's window or press Ctrl-C in it. Next session, double-click again: startup is quick once installed.
Prefer a terminal, or on Linux? Open a terminal in the unzipped folder,
run npm install once, then npm run dev, and open http://localhost:3810.
The manual's installation chapter spells
this path out step by step.
Demo mode vs. normal mode. npm run demo (any OS) starts the same viewer
but never remembers a previously opened scan, so a recording or screenshot
session can never accidentally surface a real case: it is a privacy
guarantee, not a reduced feature set. The double-click starters and
npm run dev start normal mode, which restores your last opened scan. On
first launch both modes show the phantom.
If something goes wrong
macOS refuses to open Start CBCTScope.command ("Apple could not verify" or "unidentified developer"): open System Settings, click Privacy & Security, scroll down, click Open Anyway, and confirm Open in the prompt that reappears. macOS asks once; afterwards a plain double-click works. On older macOS versions the same override is a right-click on the file and Open.
Windows shows a security warning for Start CBCTScope.bat ("the publisher could not be verified", or a SmartScreen notice): choose Run (under SmartScreen, More info, then Run anyway). Windows asks this for files downloaded from the internet; the starter only installs and starts the viewer inside its own folder.
You cannot install Node.js on this computer (no administrator rights, a managed clinic PC): nodejs.org also offers a standalone binary that needs no installer, and CBCTScope itself writes nothing outside its own folder and
~/.cbctscope. Whether either may run on a managed computer is your IT department's call; ask them rather than working around a policy.npm: command not found(or'npm' is not recognized): Node.js is not installed yet, or the terminal was opened before the install finished. Install it, then open a new terminal window and retry.The start command warns that port 3810 is in use: the viewer is already running in another window. One is enough: just open http://localhost:3810, or stop the older one first.
"These DICOM files are compressed" when opening a scan: the viewer reads uncompressed DICOM only. Re-export from your scanner software with compression off.
Use it in the browser
The same viewer also runs as a static web build: no install, no Node.js, just
a current desktop browser. It is served at
cbctscope.com, hosted on Vercel, which builds and
publishes it from the main branch; the desktop installers come from the
GitHub releases. It opens on the built-in phantom; Open, then Open DICOMDIR or DICOM
folder or Open a single DICOM file, or a drop on the viewer, reads your
own export, with a progress panel until the volume is on screen.
The scan never leaves your computer: the browser reads the files you choose
into the tab, and the page sends them nowhere. The site is static files, and its Content Security Policy forbids the page to
contact any host but its own and the model providers of the in-page agent
below, which receive nothing unless you switch the agent on. Marks are kept in the browser's own
storage and travel as a marks file (Save marks to file / Load marks from
file). The browser must hold the whole volume in memory, so a large volume
shows a notice first; the repository viewer holds it in a tab as well and shows no notice. AI-agent control (MCP)
needs the local app or the desktop app. Build it yourself with npm run build:web
(writes out/). The user manual is served by the viewer itself, on the site and in the
app, from Help.
The in-page agent is the browser's answer to MCP, and it is off until you switch it on. Type a request ("show the left TMJ in sagittal, bone window") in the agent panel, and a model on your own account (Anthropic, OpenAI, Google Gemini, or a model server on your own computer such as llama-server, Ollama or LM Studio) moves the viewer with the same fifteen navigation verbs. Your API key stays in the tab's memory and is forgotten when the tab closes; the snapshots the agent asks for go to the provider you chose and nowhere else; the agent only moves the camera and nothing it says is a finding. The desktop app has the same panel next to its MCP bridge (manual section 9.6).
The desktop app is the same web build in its own window with the agent bridge built in, so no Node.js and no browser tab: download the macOS dmg (signed and notarised by Apple) or the Windows installer (unsigned, so Windows shows its SmartScreen notice once) from the releases page. On the website, the header's Download the desktop app menu offers the same files as direct downloads, your own system first. The MCP setup below works with it unchanged (manual section 2.6).
Opening your own scan
Click Open in the viewer header and pick a CBCT export: a folder of DICOM slices, a DICOMDIR tree, a single multiframe DICOM file, or one slice of a series (the whole series opens). Any export your scanner software produces as uncompressed DICOM works.
The scan is read in place from the folder you picked. Nothing is copied, converted, or uploaded, and the full folder path never reaches the browser: the header shows a short label built from the folder and file names, so name your export folders the way you want to see them.
The file dialog is native on macOS and Windows out of the box. On Linux it uses the desktop's standard dialog helper (zenity or kdialog; most distributions ship one). If neither is installed, point the running viewer at an export folder with one command in a second terminal (forward slashes work in Windows paths):
curl -X POST http://localhost:3810/api/cbct/source -H "Content-Type: application/json" -d "{\"path\":\"C:/scans/patient-export\"}"Reading modes
Mode | What it does |
MPR | Orthogonal slices with crosshairs, oblique rotation, slab/MIP, measurements, 3D render with adjustable styles, clipping, and a render eraser |
Grid | Many parallel slices of a rotatable stack on one screen |
Pano | Draw the dental arch once, get a curved panoramic plus cross-sections, nerve tracing, and focal-trough control |
TMJ | Both condyles side by side in axis-corrected sections |
Reslice | A fresh slice stack along any drawn line or curve |
Ceph | A virtual cephalogram, the volume projected flat like a film |
Region | Region growing with HU presets, volume and density stats, airway analysis |
Stitch | Rigid registration and fusion of two volumes of the same subject |
Per-mode reading guides live in docs/reading-modes/.
2D radiographs
The viewer also opens single 2D radiographs (DICOM modality PX, DX, CR, or IO: panoramic, digital and computed radiography, intraoral), uncompressed 16-bit files, the same constraint as volumes. A radiograph gets a dedicated single-image view: window, gamma, invert, cursor-centered zoom, pan, fit. Gray values are display-normalized to a 12-bit scale at load (radiographs carry no HU calibration), and MONOCHROME1 files are flipped so bright is always radiopaque. The volumetric reading modes apply to volumes only. This exists so multi-reader studies can read both image classes in one instrument under identical display settings.
Exporting
The volume header has an Export menu with three client-side formats: a PNG slice
stack (a zip of one PNG per slice at the displayed window, plus a geometry meta.json),
a NIfTI volume (.nii.gz, HU values, for Slicer, ITK-SNAP, nibabel, MONAI), and an STL
surface mesh (marching cubes at a chosen HU threshold, honoring the 3D crop box, for
printing or CAD). Everything is computed in the browser and saved to this machine's
Downloads: nothing is uploaded, the same local-first constraint as reading.
Privacy
The safe behavior is the only behavior the software has:
Scans are read in place from the folder you pick. Nothing is copied or uploaded anywhere; there is no account, and no instance receives scans: the site is static files. The browser reads the scan into the tab, and the scan stays there unless you switch on the in-page agent, which sends the snapshots it asks for to the model provider you chose.
Everything, including the AI-agent connection, stays on this machine. In technical terms: the server binds to 127.0.0.1 and rejects any request whose Host or Origin is not loopback, so neither another device on your network nor a hostile web page in another tab (DNS rebinding, cross-site POSTs) can reach it.
No absolute file paths reach the browser or the agent transcript. The UI shows a display label derived from the last folder and file names of what you opened, plus technical geometry only.
The only writes are an app-data folder (
~/.cbctscope) holding your annotation sidecars (labels, world-mm coordinates, HU statistics) and a pointer to the last source you opened, so the viewer can restore it. Never pixels. Demo mode writes neither.
AI-agent control (MCP)

Above: an agent driving the viewer. Slice navigation, the window presets, every mode switch, and the closing view reset are MCP verbs sent by the agent; the 3D orbit, the arch, the focal trough and the cephalogram contrast are the viewer's own one-click tools. The volume is the same real CBCT as the stills above.
If you use an AI assistant that speaks MCP (Claude Code, Claude Desktop, and others), it can drive the viewer for you: open a scan (or let you pick one in the native dialog, so it never sees the path), switch reading modes, set the window, step through slices, walk the panoramic arch, bookmark and restore saved views, take snapshots to look at. It cannot do more than that, by design: no tool returns a finding or an interpretation, and no tool executes code, so the agent can never become the reader. A view the agent saves is badged agent in the viewer, so a bookmark is never mistaken for yours. A snapshot the agent asks for goes to the agent's model: under a cloud host it leaves this computer as part of that chat, under a local host it does not.
Every verb declares what it does (read-only or not, never destructive), returns
structured state, and comes with read-only resources and four ready-made
reading choreographies as prompts; docs/mcp.md has the full contract. For
Claude Desktop, each release also carries a one-click cbctscope-<version>.mcpb
bundle of the server, and the repository ships an Agent Skill in
skills/cbctscope-reading/.
Start the viewer, open it in a browser, then register the MCP server with your agent host:
{
"mcpServers": {
"cbctscope": {
"command": "node",
"args": ["/absolute/path/to/cbctscope/mcp/server.mjs"]
}
}
}The agent gets fifteen verbs, all navigation or visualization: open_scan, pick_scan, list_volumes, select_volume, viewer_state, set_view_mode, set_window_level, set_3d_style, navigate_slice, navigate_arch, list_views, goto_view, save_view, snapshot, reset_view. Details and the full contract: docs/mcp.md.
CBCTScope is one of a pair of agent-native instruments built on the same principle (bounded verbs for the agent, commit rights for the human); the other is Quoin, a text editor for macOS where any agent may edit the buffer and only the human holds the save button. Both at rezamotaghi.com.
Why this exists
CBCTScope was built on one rule: the agent moves the camera, the human reads. The field has since arrived at the same place. The first benchmark of agents operating a radiology viewer (ABRA, May 2026) found ten current models executing viewer actions at 89 percent or better while reaching correct outcomes on only 0 to 25 percent of real annotation tasks, and located the bottleneck in perception, not in tool orchestration. A navigation-only agent surface is the strong half of that result, by design. Neighboring projects exist and deserve credit: dicom-mcp and the Flux Inc. DICOM MCP server query PACS metadata over MCP, dicom-viewer-mcp-app prototypes MCP-driven DICOM slice rendering, and mcp-slicer drives 3D Slicer through raw Python execution. PrismaMRI is a general MRI and CT viewer whose Claude Desktop MCP bridge, since 31/05/26, lets the agent navigate, set the window, measure, pin findings graded by severity, and write a structured description of the study that the user can export as a PDF report. The Agent module of Slicer Automated Dental Tools, since 20/07/26, has a local model (Qwen3-8B through Ollama) pick, configure, and run one of the extension's CBCT and intraoral-scan analysis tools, such as segmentation, landmarking, or orientation, from a natural-language request inside 3D Slicer. None of these six is a purpose-built CBCT reading environment, and none constrains the agent to a small set of clinical navigation verbs: each gives the agent something other than the camera, PACS queries, code execution, analysis tools, or the findings themselves. CBCTScope does both: a dental and maxillofacial reading workflow, plus an agent surface that is deliberately navigation-only.
Development
npm run dev # dev server on :3810
npm run typecheck # tsc
npm run lint # eslint
npm test # vitest: geometry, reformats, registration, region growing, HU windowing, phantom, manual, citation and posture drift
npm run build # production buildBuilt on Cornerstone3D and Next.js. All rendering styles, reading workflows, and reconstruction math are original, from-scratch implementations informed by clinical reading practice; the project contains no vendor code, assets, or documentation.
Author
Built by Dr. Reza Motaghi, board-certified oral and maxillofacial radiologist, 100,000+ studies reported, who trains imaging models on his own reads and tests them on cases they have never seen. CBCTScope is the viewer I read CBCT with. If you use it in research, please cite it (see CITATION.cff).
License
AGPL-3.0-or-later. Free to use, study, and modify. If you build a product or service on this code, its source must be open under the same terms.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
PDF and image tools for AI agents: compress, merge, split, OCR, sign. Runs locally, never uploads.
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
PDF, image, video, OCR, screenshot, SQL, QR and text tools for agents. No API key, no signup.
Hosted browser for agents: open pages, search the web, extract data, read PDFs and take screenshots.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceExposes DICOM operations as tools for AI agents to interact with medical imaging systems and PACS. It enables searching for studies and series, querying instance metadata, and retrieving pixel data from DICOM entities.-
- AlicenseNot gradedqualityDmaintenanceProvides AI-powered medical image analysis tools for LLM agents, enabling tasks such as X-ray classification, interactive segmentation, and visual question answering. It supports multi-step diagnostic reasoning and clinical workflows through a suite of specialized medical AI models.1MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI coding agents to perform end-to-end medical image analysis locally, from inspecting CT/MRI/X-ray studies and running containerized segmentation models like MedSAM2, TotalSegmentator, and HD-BET to refining masks, measuring volumes, and writing reports. It pairs preset or custom guideline protocols with a code-customizable viewer that renders images back to agents as annotated PNGs and to humans as interactive HTML.2Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables a connected LLM to inspect brain MRI slices as crops and montages with coordinates, compare a region against its mirrored hemisphere, and review the pipeline's candidate hotspots. It lets the model write prioritised "area to review" annotations that appear live in the browser viewer, and move or read the clinician's crosshair position.MIT