Create agents
out of thin air.
TensorSketch is an open platform for building AI agents that are code-first and visually editable. Write Python, or sketch on a canvas — it's the same graph from two angles.
pip install tensorsketch-core
One idea, three ways in.
The framework, the canvas, and the visualizations are all open source — and all the same story: your work is always just code.
tensorsketch-core
A durable, code-first agentic framework. Tools, three model providers, multi-agent orchestration, tracing, and evaluation — all in plain, typed Python.
Build agents in code → CanvasTensorStudio
Draw an agent on a canvas and get the equivalent Python — losslessly, in your own style. Or open existing code and see it as a living diagram.
Sketch an agent → PlaygroundTransformer Visualizer
A 3D, equation-driven view of a transformer's forward pass. Edit the equations and watch the tensors change shape in real time.
Launch the viz →Why it's built this way
Four commitments keep TensorSketch easy to start and hard to outgrow.
Code is the source of truth
The canvas is a lossless projection of your Python — never a separate format you have to keep in sync.
Durable by default
A BSP runtime with checkpoints, crash-resume, and exactly-once side effects. Long agents survive restarts.
Everything is a plugin
Providers, tools, storage, and protocols sit behind small interfaces. Start tiny; swap anything.
Bring your own everything
No vendor lock-in, no forced database, no UI you can't leave. The core depends on almost nothing.
Open source, end to end.
TensorSketch is an open project under the tensorsketch-ai organization. The framework, the studio, and these visualizations are all public — no accounts, no lock-in. Read it, fork it, build on it.