Preface

The common practice for 3D modeling in DeepSeek Harness (DSH) involves having the agent write a bpy script and then manually running it in Blender. Steps like asset import, multi-view preview, goal-based validation, and GLB export are often scattered outside the conversation, making it difficult to establish a reproducible workflow.

Here, we introduce dsh-blender (SkillHub directory name cheshirejcat/blender), maintained by CheshireJCat. It integrates modeling, reconstruction, rendering, validation, and export into an installable plugin: registering 1 orchestration Skill, 29 domain Skills, 26 analysis/validation Helpers, and 13 workspace-level Blender tools. Each operation launches a local Blender background process, requiring no Blender plugins or control ports.

What is This

dsh-blender is an npm package dsh-blender (current version v0.2.1), classified as a workflow plugin. It uses create-3d-model-skill as a blueprint, enabling DSH agents to complete the entire 3D production pipeline—from rough modeling to export—within the session workspace.

Maintained by CheshireJCat, the source code is hosted on GitHub (approximately 19 stars). The directory page can be found on SkillHub. SkillHub is an independent DSH plugin community site, with no official affiliation to DeepSeek or High-Flyer.

Core Features

Blender Tools (13)

The plugin registers the following executable tools for DSH, all completed via a local Blender background process:

  • blender_status: Checks Blender, analysis Python, Skills, Helpers, and the current workspace status.
  • blender_scene_info / blender_object_info: Read-only views of scene and object-level geometry, materials, UVs, constraints, shape keys, and animations.
  • blender_import: Imports GLB/glTF, FBX, OBJ, STL, USD, PLY, or DAE files and saves them as versioned .blend files.
  • blender_python: Executes small, vetted segments of bpy/bmesh code and saves versioned .blend files.
  • blender_preview: Generates temporary lighting previews for isometric, front, back, left, right, top, and bottom views without relying on existing cameras, without modifying the source file.
  • blender_render / blender_render_frames: Outputs PNG/JPEG images for single frames or representative animation frames for visual validation via read_image.
  • blender_export: Exports GLB/glTF, FBX, OBJ, STL, USD, or PLY files.
  • blender_validate_scene / blender_validate_export: Audits scenes against general, web, 3D printing, or animation goals and re-reads exported files in a clean process.
  • blender_helper_catalog / blender_helper_run: Browses and executes all 26 reference image, wireframe, outline, multi-view, UV, texture, appearance, repair, and animation QA Helpers.

Skill and Helper Stack

Among the 30 Skills, create-3d-model handles the overall orchestration; others like blender-modeling, reference-to-3d, wireframe-to-3d, blender-animation, etc., can be directly loaded by DSH. The 26 Helpers provide deterministic reference analysis and verification capabilities, working with the aforementioned tools to complete blockout, refinement, materials, lighting, cameras, animation, reconstruction, and asset handoff.

Prerequisites

Ensure the following dependencies are available before installation:

  • DeepSeek Harness 0.1.0-rc.6
  • Node.js 20+
  • Blender 4.3+ (implementations are already adapted for Blender 5.x APIs); blender must be in the PATH, or an absolute path must be configured in cordis.patch.yml
  • Python 3.10+ if reference images, wireframes, or visual measurement Helpers are needed.

Installation and Activation

The official installation steps are provided in the SkillHub and README. First, install the plugin, then execute the analysis environment initialization (optional; see the note below):

npx @deepseek-ai/dsh plugin --profile web add dsh-blender
npx @deepseek-ai/dsh plugin --profile web exec dsh-blender-setup
npx @deepseek-ai/dsh --profile web --dump-config
npx @deepseek-ai/dsh web

dsh-blender-setup creates a private .venv within the installed plugin directory and installs OpenCV, NumPy, Pillow, and SciPy. The second step can be skipped if reference images, wireframes, multi-view, texture, or animation analysis Helpers are not needed.

To pin a specific GitHub Release version:

npx @deepseek-ai/dsh plugin --profile web add github:CheshireJCat/blender#v0.2.1

For development from source, run in the repository root directory:

pnpm install
pnpm setup:analysis
npx @deepseek-ai/dsh plugin --profile web add .

Common Configuration Options

Default configuration is located in cordis.patch.yml and can be adjusted as needed:

config:
  blenderExecutable: blender
  analysisPythonExecutable: ''
  timeoutMs: 180000
  helperTimeoutMs: 120000
  maxOutputChars: 20000
  restrictToWorkspace: true
  enablePython: true
  enableHelpers: true
  enableMaintenanceHelpers: false
  registerSkill: true
  registerModuleSkills: true

When restrictToWorkspace: true, inputs and outputs must remain within the current session workspace. Setting enablePython: false disables blender_python, while retaining checks, rendering, and export. Setting registerModuleSkills: false registers only the orchestration Skill, not the 29 domain Skills.

Typical Usage

After installing and starting the Web UI, open http://127.0.0.1:3080, create a new session with the modeling directory as the workspace, and send the following task to DSH:

Use create-3d-model to create a low-poly desk lamp.
First call blender_status, save and render the rough model and final version separately for validation,
and finally export GLB while keeping versioned .blend source files.

For a one-off headless verification without using the Web UI:

npx @deepseek-ai/dsh plugin --profile headless add .
npx @deepseek-ai/dsh --profile headless \
  "Use create-3d-model to create a low-poly blue cube under artifacts, render a preview, and export GLB."

The example output in the README is a camera-less lamp preview from a real installation smoke test.

Applicable Scenarios and Notes

This is suitable for developers and agent workflow builders who need to complete a full Blender production pipeline within DSH sessions: from reference image or wireframe reconstruction, blockout, materials, lighting, animation validation, and multi-format export—without manually switching tools outside the conversation.

Please note the following before use:

  1. The plugin runs with the permissions of the current dsh process; review the source code and MIT license before installation to ensure the security boundaries meet expectations.
  2. blender_python has the same permissions as running local Python within Blender and should only be used for trusted modeling code. The plugin disables automatic execution of embedded scripts in .blend files by default, restricts paths to the workspace, defaults to rejecting overwrites of existing files, and uses a separate background process for each invocation.
  3. No BlenderMCP or additional Blender plugins are required; each operation directly launches a local Blender background process.

Links

dsh-blender consolidates the fragmented script calls between DSH and Blender into an orchestration-ready production workflow: checking, importing, modeling, previewing, rendering, validating, and exporting are all completed within a unified set of tools and Skill stacks. It is ideal for integrating 3D asset production into an agent’s closed-loop validation session.