Your data stays on your machine
Conversations, canvas state and run history are saved to a SQLite file on your own disk. No account and no cloud sync.
./data/geochat-desktop.sqlite
GivenConstructConclude
GeoChat puts an AI assistant and GeoGebra in one desktop math tool. Enter a problem and it plans the construction, writes GeoGebra commands into the canvas, and explains the reasoning behind each step.
macOS and WindowsApache-2.0, open sourceYour data stays local
Draw directly in DeepSeek Harness
Give Harness a GeoGebra canvas. Your agent uses native tools to construct, read back the scene and inspect screenshots. Drag objects and sliders on that same canvas, without installing GeoChat Desktop or connecting a separate MCP service.
Construct, then check the result
- 20 native tools, 55 mathematical skills
- Open a session canvas, search commands, construct objects, inspect results and control animations. Applet assets ship in the package, without an external CDN.
- Let the agent see the actual canvas
- geochat_capture_canvas returns directly viewable image content. Frame sampling helps inspect changes during animation, beyond numeric object values.
- Read back scenes and text layout
- Read the viewport, styles and partial dependencies; check container overlap and bounds for named text and bounded text lists in the current frame. Ordinary object labels and all animation instants are not covered; screenshots remain necessary.
Install in Harness
Harness 0.2.0-rc.2 · Node.js 24+ · Web profile (not the Desktop carrier)
dsh plugin --profile web add https://github.com/tiwe0/GeoChat/releases/download/dsh-plugin-v0.1.5/dsh-geochat-plugin-0.1.5.tgzRestart the matching Harness profile after installation, then read the geochat skill. The plugin uses your configured Harness model; no separate plugin key is needed.
Try: Draw an animated solar system on the GeoChat canvas, then capture it to check labels and layout.
Release notes and SHA-256 checksumsEach session has its own in-memory canvas. Refreshing, closing the canvas or restarting loses the construction; save anything you need first. Bundled offline assets do not mean no network: Harness manages the native Remote connection.
Plugin-owned code is Apache-2.0. Bundled GeoGebra assets have non-commercial license restrictions; commercial use or distribution requires separate permission. GeoGebra licensing
What GeoChat does
These are the features currently supported in the desktop app.
What is given or stated
Bring your own model key
Add your model provider's API key in settings. Requests go directly from your machine to that provider, without passing through GeoChat.
OpenAI · Anthropic · Google · DeepSeek · Alibaba · OpenRouter
The construction in progress
Write commands into GeoGebra
GeoChat writes real commands into the embedded GeoGebra canvas. The resulting figure can be dragged, measured, and edited further.
O = Intersect(m_1, m_2)
Plane and space, one canvas
The GeoGebra runtime ships inside the app. Plane geometry, coordinate geometry, function graphs and solid geometry all use the same workspace.
What is concluded
See every step
The construction plan, tool calls and reasoning are shown in order so you can check them one by one.
Problem bank and export
Import a local problem bank to work through sets in bulk, and export any canvas as a .ggb file that a student or colleague can open in their own GeoGebra.
*.ggb
How a problem is completed on the canvas
Read, plan, write commands, explain. The desktop app handles all four stages; only model inference needs the network.
Given triangle ABC, construct its circumcircle.
A = (-3, -1)B = (3, -1)C = (1, 3)m_1 = PerpendicularBisector(A, B)m_2 = PerpendicularBisector(B, C)O = Intersect(m_1, m_2)k = Circle(O, A)SetCaption(O, "Circumcenter")Read
Identify the given conditions, constraints and goal.
Plan
Choose the construction order and the geometric objects needed at each step.
Compile
Convert the construction plan into GeoGebra commands and write them into the canvas one by one.
Explain
Explain why these objects produce the requested result.
Full demo
From typed problem to finished figure, shown at real speed.
The desktop app
The canvas is the main view. The chat panel can be collapsed or moved, and model settings live in the app's local settings page.
Questions
- Does it need an internet connection?
- The canvas, database and other local features work offline. Only model inference needs the network; requests go directly from your machine to the provider you configured.
- Does it cost anything?
- The app is free and open source under Apache-2.0. You pay your model provider for your own usage; that billing sits entirely with them, and we neither see it nor take a cut.
- Do my problems get uploaded to you?
- No. We do not run a server that could receive them. Your problem is sent as a prompt to the model provider you configured, and that data is governed by their privacy policy.
- Which model providers are supported?
- OpenAI, Anthropic, Google, DeepSeek, Alibaba Cloud and OpenRouter are configured out of the box, and you can add compatible custom providers and models in settings.
- Is there a Linux build?
- The packaging pipeline currently produces macOS and Windows installers only. Linux works from source, and the repository has full build instructions.
- Are the installers code-signed?
- The macOS installer is signed with Developer ID and notarized by Apple, so it opens normally. Authenticode signing has not been verified for the Windows installer, so SmartScreen may still appear on first run; the download page explains what to expect.
Install it, give it a problem
Open source, free, no sign-up. All you need to bring is an API key from a model provider.