loading
loading
DeepSeek Harness Lets AI Build Full 3D Scenes Autonomously in Blender A new open-source plugin wires an AI agent directly into Blender over TCP, exposing 16 tools and 186 ops for scene editing, rendering, and verification. - sixtysevenlf released dsh-blender-plugin v2.1.0, an open-source bridge letting AI agents drive Blender over direct TCP. - Exposes 16 model-facing tools and a plan channel of 28 families and 186 ops, frozen by a capability lock in tests. - Skips MCP in favor of two local ports (9876 addon, 9877 backend) with a 15-second watchdog and write-lease protection. - Includes a contract layer with destructive guards, three-state verdicts (supported, refuted, unresolved) and md5-stamped evidence ledgers. - Supports WSL, native Linux, Windows, and macOS with Blender 4.x and 5.x in GUI mode, licensed BSD-3-Clause. - Pairs with a separate blender-modeling skill covering acceptance criteria and multi-agent modeling workflow. Developer sixtysevenlf has released version 2.1.0 of the GitHub repository, described as the first stable release of an open-source bridge between large language models and Blender. The BSD-3-Clause project gives an agent direct control of a live Blender session through local TCP connections, while a separate modeling skill defines the workflow and acceptance criteria. For developers building autonomous 3D pipelines, the main contribution is a bounded API that combines viewport capture, scene editing, search loops, rendering, rollback, and verification. The model-facing surface stays fixed at 16 tools, while the planning layer exposes 186 operations across 28 families. Two loopback ports carry the session The add-on and backend communicate through two loopback endpoints: the Blender add-on uses 127.0.0.1:9876, and the backend uses 127.0.0.1:9877. The project uses its own socket protocol and omits a Model Context Protocol server, packaging the integration as a profile bundle for DSH, or DeepSeek Harness. The channel supports viewport capture, Python execution, time-window monitoring, inner search loops, render profiling, safe decimation, headless processing, batch jobs, presets, rollback, and raw add-on commands. Because the agent receives frames and modifies the scene through one connection, it can inspect an edit and verify the result within the same session. Sixteen calls keep the surface bounded The release documents seven core calls in detail, with latency figures measured on the author’s machine. These figures provide implementation reference points, and actual performance will depend on hardware, scene complexity, Blender settings, and renderer configuration. This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
document fetched 2026-10-24 02:55