Inspiration
AI agents can manipulate creative software, but their work is often opaque: the user provides a prompt, waits, and sees only the final result. We wanted an agent-assisted Blender workflow where users can watch every meaningful operation, inspect intermediate results, stop the process, and retain control.
Our proof case was physical and practical: twenty animal-track stamps representing ten animals, created for a planned summer-camp activity where children can learn to find and identify animals by their tracks.
## What it does
Blender AI Agent is an installable Blender extension that gives an external AI agent a bounded, loopback-only HTTP API for working in the open Blender scene.
The agent can inspect objects, create and edit meshes, add text and modifiers, perform Boolean operations, check print readiness, and export STL files. Every mutation runs sequentially on Blender’s main thread. Blender stays open and visible, and each result appears in the viewport before the agent continues.
The extension deliberately excludes arbitrary Python execution, unrestricted Blender properties, hidden macros, and remote-network access.
## How we built it
At the beginning of Build Week, OpenCode with GPT-5.6 Sol was used to create and prove the original Blender HTTP-server workflow. That proof of concept produced all twenty physically printed animal-track stamps.
Codex was then used as an expert engineering collaborator to transform the proven concept into a distributable Blender extension. Codex helped design and implement the add-on architecture, bounded API, main-thread dispatcher, sequential execution model, visible activity UI, modeling operations, print- readiness integration, tests, isolated development harness, documentation, packaging, and release verification.
Humans made the key product and safety decisions: keeping work visible and sequential, prohibiting arbitrary code execution, restricting the service to loopback, preserving the original proof of concept separately, and using tactile educational artifacts as the real-world proof case.
## Challenges we ran into
The central challenge was balancing agent capability with safety and transparency. Complex modeling had to remain practical without accepting arbitrary scripts inside Blender or hiding an entire workflow in one request.
Independent agent testing revealed missing capabilities such as Boolean Difference and Union, topology inspection, procedural displacement, modifier removal, visibility control, and recovery. We added these as separate, bounded, inspectable operations.
We also had to protect unrelated Blender projects running on the same computer. The development harness uses isolated profiles and strict process-ownership verification so it cannot control or terminate another Blender instance.
## Accomplishments that we're proud of
- Produced twenty real, physically printed animal-track stamps.
- Reimplemented the proof of concept as an installable Blender extension.
- Preserved visible, sequential execution throughout agent workflows.
- Added bounded mesh, text, modifier, Boolean, print-check, recovery, and STL-export operations.
- Created a machine-readable
/helpendpoint so agents can discover capabilities and limits. - Tested installation and lifecycle behavior in a clean Blender profile.
- Recorded a complete installation-to-sliced-artifact workflow.
- Published a downloadable, checksum-verified beta release.
## What we learned
AI-assisted creative work is more trustworthy when the application remains the visible workspace. Users should not have to surrender observability to gain automation.
We also learned that agents can calculate complex deterministic geometry externally and submit bounded vertex and face data without requiring arbitrary code execution inside Blender.
Finally, independent agent testing was extremely valuable: it exposed practical workflow gaps much faster than designing the API from assumptions alone.
## What's next for Blender AI Agent
The Build Week beta is verified on Blender 5.2.0 LTS under Ubuntu 24.04. Future work includes testing other operating systems and Blender versions, improving first-run guidance, expanding safe modeling capabilities, and gathering feedback from educators, makers, and agent developers.
The animal-track stamps are planned for use at summer camp. That educational activity has not yet occurred.
Log in or sign up for Devpost to join the conversation.