use-blender

Why it's ranked here

A narrow, carefully bounded piece of plumbing. The whole project is about 500 lines of Python across four runtime files, MIT licensed, with Blender pinned to an exact distribution package on a digest-pinned base image. The input side validates every action batch against the screen bounds and a 15-second time budget before anything reaches the display, script execution is off by default, and a failed request whose outcome is unknown locks the session until restart rather than guessing. Against that: an early project with no tests, no prebuilt image, and only one platform tested by its authors.

What's good

The safety defaults are specific. The published port binds to the host's loopback address, an optional bearer token guards every endpoint, and file downloads refuse any path that resolves outside the workspace. Only one mutating request runs at a time, a second gets a clear busy error, and request bodies are capped at 64KB. Typed text goes in through the clipboard, so non-English characters arrive intact regardless of keyboard layout. The container ships a health check, runs as an unprivileged user, and the compose file already caps it at two CPUs and 4GB of memory.

Tradeoffs

The README says it is tested only on Linux ARM64 through OrbStack and that AMD64 has not been validated. There is no registry image: every user builds it locally from the repository. The repository tree has no test directory and no CI configuration. Python scripts run on Blender's main thread, and the README warns that a timeout neither stops nor undoes them, so a long script can freeze the session. The README also notes Blender exposes its own Python console, so the token does not make untrusted clients safe. There is no video stream; agents work from still screenshots only.

How to use it well

Fits a team building or evaluating computer-use agents that need a real, visually complex desktop application to practise on, or anyone who wants an agent to produce Blender files without a GPU. Keep batches short, move the pointer over the right editor before sending shortcuts, and take a fresh screenshot after each batch. Turn on scripting when precision matters more than testing the agent's clicking. Mount a named volume so saved files survive container replacement. It does not include an agent, a planner or live video; bring your own loop.

Observed

License

MIT for the project's code; Blender and system dependencies keep their own licenses

Language

Python

Packaging

Dockerfile and compose.yaml, built locally; no registry image

Interface

HTTP API: health, screenshot, state, download, actions, and opt-in Python execution

Hardware

CPU-only software rendering, no GPU required

Platform tested

Linux ARM64 via OrbStack per README; AMD64 not validated

Authentication

Optional bearer token via API_TOKEN

Tests

No test directory or CI configuration in the repository tree