Preface¶

DeepSeek Harness (command name dsh) runs coding agents locally. It is proficient with text, code, and file read/write operations, but when it comes to 3D assets, the common workflow becomes: humans build models in Blender, then export and submit the GLB back to the project. The agent can neither see the scene topology, correct outlines using reference images, nor perform a clean process validation before export.

DeepSeek officially designed Harness with the principle of “everything is a plugin”: models, tools, skills, sessions, sandboxes, loops, and UI are all replaceable plugins without modifying the Harness source code. The community directory deepseek-harness-plugin.com hosts many third-party extensions. It is an independent website and has no official affiliation with DeepSeek / Fang Intelligence, and should not be treated as an official app store.

The blender plugin is listed under the “Tools & Capabilities” category in this directory. Instead of connecting DSH to a remote MCP or Blender plugin port, it launches a local Blender background process, unifying modeling, reference image reconstruction, rendering validation, verification, and export into a single pipeline. This article is organized after cross-referencing the directory details page, the repository’s README.md / README.zh-CN.md, package.json, CHANGELOG.md, NOTICE, and official DeepSeek Harness documentation.

What is this¶

The blender plugin is maintained by CheshireJCat, with the npm package name dsh-blender. The current repository version is 0.2.1 (2026-08-14), licensed under MIT. The community directory page marks it as primarily written in Python; the repository itself is a Node.js plugin, while the actual scene modification is handled by a Blender background process launched on each call. Reference image and visual metric helpers run OpenCV, NumPy, Pillow, and SciPy using Python 3.10+.

It does not solve the problem of “describing a model in a conversation”, but rather tasks such as:
- Create low-poly / production-grade modeling based on text or reference images
- Reconstruct exportable meshes from wireframes, orthographic views, and texture atlases
- Apply materials, lights, and cameras to scenes, and generate render evidence readable by read_image
- Audit scenes for web, 3D printing, or animation targets, then export in formats like GLB / FBX
- Continue editing existing .blend or portable assets instead of starting from a blank scene every time

The plugin documentation clearly states that this is not an official product of DeepSeek AI or the Blender Foundation. The modeling skill is adapted from create-3d-model-skill, and the module recipes and helper scripts retain MIT licensed content from the upstream RobLe3/cc-blender-skill. Details can be found in the repository’s NOTICE and skills/create-3d-model/references/upstream.md.

When checking the repository page, GitHub shows 16 stars, while the community directory page marks 8 stars; the repository star count shall prevail.

Core Features¶

The plugin registers 1 master orchestration Skill (create-3d-model), 29 domain Skills that can be loaded directly, 26 controlled Helpers, and 13 workspace-scoped Blender tools. It does not require installing BlenderMCP or opening control ports; each tool call uses an isolated Blender background process, with temporary scripts and JSON cleaned up after execution.

13 Runtime Tools¶

Per the repository’s README.zh-CN.md and references/dsh-integration.md, the tools directly callable from the model side are as follows.

The recommended execution order is written in the integration documentation: first blender_status; if a .blend file already exists, check it first, otherwise use blender_import first; execute blender_python in small steps; use preview/render with read_image for validation; run blender_validate_export after exporting. When the user does not specify a format, .glb is delivered by default, and .blend is treated as an editable project source file.

How the 30 Skills Are Divided¶

The master orchestration entry is create-3d-model. The remaining 29 modules are grouped in capability-map.md, and can be loaded using DSH’s skill tool with their kebab-case names when needed. Do not dump all documentation into the context at once.
- Production Control: blender-pro-workflow, blender-skill-harmonizer, quality-refinement-autoloop, mascot-logo-reconstruction
- Blender Domains: blender-modeling, blender-materials, blender-lighting, blender-cameras, blender-rendering, blender-animation, blender-export, blender-uv-texturing, wireframe-to-3d
- Reference Image Locked Reconstruction: reference-to-3d, reference-analysis-validator, source-part-segmentation, contour-to-mesh, orthographic-registration, multiview-constraint-solver, atlas-uv-fitting, etc.
- Animation QA: texture-state-animation, orbital-hud-motion, animation-quality-gate

For generating models from text or a single image, the recommended path is: blender-pro-workflow → modeling → camera / lighting / materials → rendering → export when needed. Reconstruction modules should only be loaded when reference images need to be aligned to outlines, and do not mix “looks like” and “1:1 fit to source image” in the same task.

Analysis Helpers and Blender Versions¶

26 Helpers are exposed via a catalog and runner. The CHANGELOG notes that 23 of them are modeling-related and enabled by default, while 3 maintenance Helpers require enableMaintenanceHelpers to be turned on before execution. Running pnpm setup:analysis or dsh-blender-setup will create a Git-ignored private .venv in the plugin directory and install OpenCV, NumPy, Pillow, and SciPy. Skip this step only if you will not perform reference image, wireframe, multi-view, texture, or animation analysis at all.

Environment requirements from the README:
- DeepSeek Harness 0.1.0-rc.6
- Node.js 20+
- Blender 4.3+ available in PATH as blender, or specify the absolute path in the configuration
- The current implementation is adapted for Blender 5.x API; the runtime version should be confirmed via in-process bpy.app.version, not just the local installer name

Installation and Activation¶

The installation command given on the community directory page is:

dsh plugin add github:CheshireJCat/blender

The directory page also reminds users to pin the commit hash for reproducible installations:

dsh plugin add github:CheshireJCat/blender#commit

Replace #commit with the actual commit hash. The plugin runs with the permissions of the current dsh process, and may execute code during installation. You should review the source code and license before installing.

The repository README additionally provides an npm installation method for the web profile, with analysis environment initialization:

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

You can also pin to the GitHub Release v0.2.1:

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

For development from source, run dsh plugin --profile web add . at the repository root, then dsh --profile web --dump-config to confirm the bundle layer is attached, and restart the Web UI. The default interface address is http://127.0.0.1:3080. When creating a new session, point the workspace to your modeling directory instead of an arbitrary empty folder.

Typical Usage¶

The session example given in the repository’s Chinese README is: open the Web UI, set the workspace to the modeling directory, then tell dsh:

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

A one-time headless validation can be written as follows (at the plugin repository root):

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 integration documentation also specifies path conventions: persistent artifacts should be placed in directories like artifacts/<task-name>/; use new versioned .blend files for each major stage, such as lamp-blockout-v001.blend, lamp-final-v002.blend; do not write deliverables to /tmp; overwrite existing files is denied by default unless explicitly permitted.

Self-test commands for developers are also in the README:

pnpm install
pnpm setup:analysis
pnpm validate
pnpm test
pnpm test:blender

test:blender will launch Blender for real, covering blend file creation and import, object inspection, scene auditing, preview and official rendering, GLB export, and isolated process re-import validation.

Configuration Options¶

The default configuration is in cordis.patch.yml:

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

The meanings of several switches are based on the Chinese README:
- restrictToWorkspace: true: Restrict input and output to the current session workspace, and check the real target of symbolic links
- enablePython: false: Disable the most powerful and high-risk blender_python tool, while inspection, rendering, and export remain available
- enableHelpers: false: Disable the Helper catalog and execution tools, while basic production tools are unaffected
- enableMaintenanceHelpers: false: Disable skill publishing/cleanup maintenance scripts by default
- registerModuleSkills: false: Only register the master orchestration Skill, not the 29 domain Skills
- registerSkill: false: Only register tools, do not contribute any Blender Skills to the DSH Skill catalog

If Blender is not in your PATH, change blenderExecutable to the absolute path. You can also point analysisPythonExecutable to an existing Python environment for the analysis tools.

Applicable Scenarios and Notes¶

This plugin is suitable for situations where:
- You already have Blender 4.3+ installed locally, and want dsh to directly produce exportable 3D assets in the session workspace
- You need to continue modeling from text, reference images, wireframes, or existing assets, and want to use renderings and scene audits for validation instead of just checking file sizes

Unsupported expectations should also be clarified:
- It will not replace the detailed sculpting workflow in the interactive Blender viewport
- By default, online asset or generation services such as Poly Haven, Sketchfab, Hyper3D, and Hunyuan3D are disabled unless explicitly requested by the user
- quality-refinement-autoloop should not modify Skills, documentation, or Git status during ordinary modeling tasks

The security boundaries are based on the repository README: blender_python has the same permissions as running local Python in Blender, and should only process trusted modeling code. Do not use it to download resources, read credentials, install packages, launch unrelated processes, or modify Blender global preferences. By default, the plugin disables automatic execution of embedded .blend scripts, restricts paths to the workspace, refuses to overwrite existing files, and uses isolated background processes for each call. Even so, the plugin runs with the permissions of the current dsh process. You should read the source code and MIT license before installing, and pin commits or release tags for production environments.

Summary¶

The blender plugin (npm package dsh-blender) treats Blender as a local production backend for DSH: 13 tools handle inspection, scene modification, rendering, and export; 30 Skills handle orchestration and domain specialization; 26 Helpers handle reference image processing and validation. For users already using DeepSeek Harness who need to keep 3D assets within their workspace, this pipeline is far more complete than the workflow of “describing the model verbally, then manually building it in Blender”.

Directory page: https://deepseek-harness-plugin.com/zh-CN/plugins/blender/

GitHub: https://github.com/CheshireJCat/blender