CBCTScope
A CBCT viewer built by the radiologist who reads with it.
A complete, local-first CBCT viewer in eight purpose-built reading modes. The same instrument reads 2D radiographs (panoramic, intraoral, computed and digital radiography), and it is free, open source, and local-first: no account, no cloud, no upload, ever.
The viewer opens in your browser at cbctscope.com on a built-in phantom, nothing installed; your own export is read from your disk into the tab and uploaded nowhere. The desktop app is the same build in its own window; macOS signed and notarised, Windows unsigned.
No agent is required; CBCTScope is a complete standalone viewer. It is also, as far as I can find, the first purpose-built CBCT reading viewer with native AI-agent control, built on one rule: the agent moves the camera, the human reads. An agent connected over MCP opens scans, switches reading modes, sets the window, moves through slices, and takes snapshots, while the scan never leaves your computer. The field has since arrived at the same place: in the ABRA benchmark (May 2026), ten current models executed viewer actions at 89 percent or better and reached 0 to 25 percent on outcomes, the bottleneck in perception, not tool use.
Board-certified specialist in oral and maxillofacial radiology · 100,000+ studies reported
- Reading modes
- 8 purpose-built
- 2D radiographs
- Panoramic · intraoral · CR · DX
- Standalone
- No agent required
- Agent surface
- 15 navigation verbs, over MCP or in the page
- Privacy
- Read on your machine · no uploads · no account
- License
- AGPL-3.0, open source
- Runs
- In the browser · desktop app (macOS, Windows) · from the repository (macOS, Windows, Linux)
- Current release
- v2.1.1, November 15, 2026
Latest release: v2.1.1 (November 15, 2026) Release notes
- 01orthogonal MPR with 3D rendering
- 02a multi-slice grid
- 03curved panoramic with nerve tracing
- 04TMJ
- 05freehand reslice
- 06virtual cephalogram
- 07region growing with airway analysis
- 08two-volume stitching
The reconstruction math is original, implemented from scratch and informed by reading practice, not lifted from any vendor.
Four of the eight reading modes, on the same real CBCT as the recording. Click any still for full resolution.
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, and it is not for clinical decision-making.
Agent-native imaging
An agent-native viewer exposes clinical navigation verbs, not code execution. CBCTScope’s agent surface is fifteen verbs, all navigation or visualization: open a scan, or let the reader pick one in the native dialog so the agent never sees the path; read the viewer state; switch mode; set the window and the 3D style; move through slices, the grid and the panoramic arch; bookmark and restore saved views, each agent bookmark badged as such; snapshot; reset. Read-only resources mirror the state, four reading choreographies ship as prompts, and a one-click bundle installs the server into Claude Desktop. There is no tool that returns an interpretation. That constraint is the design: the agent can operate the instrument the way an assistant operates the viewer’s controls, and the reading itself stays with the human. Neighboring projects put agents next to imaging in other ways; the repo credits them and maps the differences.
In the browser and in the desktop app the same fifteen verbs are also reachable without an MCP host, through an agent panel inside the page. It is off on every load and sends nothing until you switch it on; then 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. Your 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; and the dialogue that switches it on states the fence: the agent only moves the camera, it is not asked for findings, and nothing it says is a finding.
The agent moves the camera. The human reads.
Privacy by construction
On every surface the scan is read from your own disk: nothing is copied or uploaded, there is no account, and no server ever receives a scan. cbctscope.com is static files; the browser reads your export into the tab, and the page’s Content Security Policy forbids it to contact any host but its own and the model providers of the in-page agent, which receive nothing unless you switch the agent on. The repository’s viewer and the desktop app bind to loopback and reject any request whose Host or Origin is not local, so neither another device on your network nor a hostile page in another browser tab can reach them.
Three ways to run it
- In the browser. cbctscope.com: nothing to install, a current desktop browser is enough. It opens on a built-in phantom, so the whole viewer can be learned with zero patient data.
- The desktop app. The same build in its own window, with the MCP bridge built in. The latest release, direct: macOS, Apple silicon, macOS, Intel (both signed and notarised by Apple), Windows installer (unsigned, so Windows asks once on first launch).
- From the repository. For Linux, a terminal, or running from source: three steps on the GitHub page.
The user manual covers all three, from installation to the keyboard reference, with a reading guide for every mode.
Agent-native instruments
One principle, two domains
Bounded verbs for the agent, commit rights for the human. CBCTScope applies it to reading CBCT volumes; Quoin applies it to text.
See Quoin



