Blender Agent Studio invokes a local Blender executable and runs Python scripts inside Blender. Review untrusted scripts before execution. The bundled MCP is intentionally limited to version checks, asset inspection, and evidence renders; it does not expose a generic arbitrary-Python tool.

Set BLENDER_EXECUTABLE only to a trusted Blender installation. Avoid running untrusted .blend files with automatic script execution enabled.

Models, exports, renders, benchmark prompts, agent traces, and evaluation output can contain private source material. They are ignored by default and should be reviewed before sharing.

The documentation builder renders repository-owned Markdown, including raw HTML. Treat changes to site/content as website code and review them before merging; the builder is not a sandbox for untrusted or visitor-submitted content. It does not fetch external Markdown or accept visitor content.

The github-pages environment must restrict deployments to the main branch, including manually dispatched workflow runs. Keep that restriction when changing repository settings.

Open a GitHub issue with a minimal reproduction. Do not include credentials, private assets, proprietary models, or sensitive agent traces.