Tsugumi
Run the analysis you sketched with AI, as it is, and check it with your own hands.
Tsugumi is a research workbench in which the specification a person draws, the thing an AI edits and the processing that actually runs are one and the same Workflow. It runs on macOS, Linux and Windows (Windows through WSL2). You draw a computation or a data pipeline as Cards and the wires between them, implement it together with an AI, and run it in environments pinned with Nix. You open the code or the result of a Card you are curious about, change a condition and try again with the inputs that were saved. Going back and forth like this, a researcher comes to understand how the analysis got where it did, and can ask the next question.
Runs on macOS, Linux and Windows (Windows through WSL2. The app needs Apple silicon and macOS 26 or later) · Installation and development builds
Learn about Tsugumi
What problem it takes on, and howGetting started
Start using it- Tsugumi's featuresFind a feature by task; Card kinds, execution rules and Nix environments
- UsageCanvas tools, running from a Card, chart updates, settings and Agent controls
- Your first CardBuild a Card that takes an input and computes, with examples in Python, Bash, C++, R, Julia, Ruby, Go, Rust and C#
- Run history and artifactsInspect and compare runs, check runtime packages and open saved artifacts
- Run a Card on a remote workerRegister a C++ remote worker and run a Code Card on its Linux host
- C# CardsC# typed ports, Tasks, NuGet dependencies and the .NET worker lifetime
- Visualization CardsEditable library/shape scaffolds, data parameters, saved figures and reports
- marimo notebook CardsCommitted notebook runs, typed inputs/outputs, saved cell results and the interactive-session boundary
- R report CardsR Markdown and Quarto report Cards, saved HTML and parameterized rendering
- Web UIThe browser workbench and the tsugumi command on Mac, Linux and Windows through WSL2
- Binary cachesAttic, Cachix and other signed binary caches, private authentication and VM-wide trust
- Using it from Claude CodeHaving Claude Code make and run Cards through the Claude Code plugin, shown in the Web UI
The ideas
Why it takes this shape- Why one Workflow for design, implementation and checkingWhy design, implementation and checking share one Workflow: answers to the three reasons for hesitating to hand research to AI (an invisible process, reproducibility, data), the measures that keep trial and error light
- Vision and applicationsApplications: learning research decisions, growing tools, handing work to the next person, and the idea of business automation
How it works
What is inside- Versions (git)Versions: the project's git, each run's version and result (git notes), the version tree and "Start from This Version"
- Security and isolationWhat is and is not protected, the four isolation boundaries (VM, project, Card, RPC), API keys, where data is kept, and known limitations
- The RPC contractThe project RPC requests and notifications (generated from protocol/tsugumi.fbs)
- GlossaryCard, Draft, Run, invocation, version and other terms
Dependencies and distribution
Pinned versions and licensesRun the demos, right here
The bundled demos. Without installing anything, you can see here how their Cards connect and how they run. Click one to open it large. Each is explained in The bundled demos.