A branching thread keeps the analysis inspectable
Data Formulator does more than turn a prompt into a chart. You load a file, connect a database, or ask its analyst to locate data, then review the proposed material before it enters the workspace. Each question produces an explanation, table, or editable visualization inside a Data Thread. You can branch from an earlier step instead of dragging stale assumptions through a single conversation. That history is the part ordinary chat-based analysis often loses.
The interface accepts CSV, TSV, Excel, JSON, screenshots, text, databases, and data platforms. Its current chart system uses Flint and the README describes more than 30 chart types in the v0.7 line. Our checkout was substantial: 689 files, about 136,350 source lines, and 28.1 MB before installation. The size reflects a TypeScript frontend, Python backend, connectors, desktop packaging, workspaces, agents, and code-execution controls rather than a thin chart wrapper.
The app can run locally, but useful answers still need a model
The easiest trial is uvx data_formulator, a pip install, or Docker Compose. A local server opens on port 5567. Windows and macOS desktop archives are also produced by CI, though the README warns that current previews are not signed or notarized. Users must consciously bypass the operating-system warning and should obtain those archives only from the project's workflow artifacts or releases.
Model access is a separate choice from installation. Data Formulator supports OpenAI, Azure, Anthropic, Ollama, and other routes through LiteLLM. Hosted providers need API keys; local Ollama use shifts the burden to model selection and hardware. Open issue 362 comes from a user who could not find a satisfactory local model even with 8 GB of VRAM and 32 GB of system memory. That report is one experience, but it shows why an offline label does not guarantee useful analysis.
What happened when we ran it
Our unprivileged Debian sandbox cloned commit 5477f0e and installed 1,050 Yarn packages in 67 seconds. The installed tree used 807 MB on disk. That is a fairly large developer environment, yet the setup completed without an error in the supplied log. The repository also includes two CI workflow files, a Dockerfile, a Compose file, and a tests directory.
The production build completed in 57 seconds. Vitest then finished in 34 seconds with 391 passed and 0 failed out of 391. No test was reported as skipped in the measurement block. Those results establish that the checked-out frontend built and its available Vitest suite passed on 3 CPUs and 8 GB of RAM. They do not measure database connector coverage, model answer quality, chart accuracy, or desktop signing.
This distinction matters because the hardest failures happen after the tool starts. An agent can select the wrong table, write a plausible transformation, or interpret a chart badly while every unit test stays green. Data Formulator makes intermediate tables and branches visible, which gives a reviewer somewhere to catch those mistakes. The 391 passing tests support the software mechanics; a human still owns the analysis.
Generated Python makes the sandbox a deployment decision
AI-generated Python runs in a sandbox. The default local backend keeps a warm subprocess and blocks file writes plus dangerous operations through audit hooks. The stronger option launches each execution in a disposable Docker container with a read-only workspace and resource limits. The documented trade is about 1 ms of overhead for local execution versus roughly 700 ms for Docker. These are project claims from the development guide, not figures from our lab run.
There is one deployment trap: the Docker execution backend does not work when Data Formulator itself runs inside its Docker image. Its child container needs host paths for bind mounts, but paths inside the parent container do not refer to the host. To use Docker isolation, run the application directly on the host. A public anonymous instance should use ephemeral workspaces and disable data connectors, custom models, and key display. Authenticated shared use needs OIDC, GitHub, or Azure identity plus properly scoped storage.
Active development comes with beta boundaries
GitHub showed 17,410 stars, 1,708 forks, and 107 combined issues and pull requests on September 27, 2026. The repository was pushed that same day, with new pull requests and dependency work updated within hours. That is healthy current activity, though the combined count should not be read as 107 bugs. The latest release is 0.8b1 from August 15, while the README still directs general users to stable v0.7.
A few open reports define the edge of the current product. Issue 398 says chart commentary can fail with an unhelpful error when a text-only model receives an image request. Issue 427 asks for semantic-layer support so governed measures and relationships do not have to be inferred from raw schemas. Data Formulator is therefore best at supervised exploration. Its 57-second build and 391 passing tests make it easy to justify a trial, but the final chart should leave the workspace only after someone checks its source, transformation, and claim.