Blender MCP bridge gives coding agents 27 scene tools
An MIT-licensed Blender extension exposes a supervised, token-gated MCP endpoint on localhost so agents can inspect, edit and export scenes directly.
A new Blender MCP bridge gives coding agents a way into Blender without screen scraping. The repository, blender-ai-copilot-bridge, was created on 12 November 2026 and is an MIT-licensed Blender extension that publishes 27 supervised scene tools on a local, authenticated MCP endpoint, so a client such as OpenCode can read a scene, add objects and export the result. GitHub showed 331 stars and 4 forks as of 2026-11-15, two days after it was created.
The endpoint and the token
The extension runs inside Blender and serves Streamable HTTP at http://127.0.0.1:9876/mcp by default. The port can be changed in the extension preferences, and the server has to be enabled there before anything connects. Each request carries an Authorization: Bearer header using a token generated when the server is turned on. The README tells users to treat that token like a password and to keep it in the client's secret store or environment.
The 27 tools fall into six groups according to the README: inspect, create, edit, capture, script and export. Three are named: get_scene_info, add_cube and export_scene. Scene inspection returns object lists, structure and details as structured context rather than screenshots, and the capture tools return viewport images that image-capable models can read.
What stays inside Blender
Supervision is the main claim. Calls arriving from outside clients land as visible, undoable operations in Blender, and cancellation and named undo apply to them as well. Approvals are split in an unusual way: external clients manage their own approval policies, while Blender keeps the undo record and makes each operation visible. So the safety net is recovery and visibility, not a gate that blocks a call before it runs.
A bundled OpenCode Portable launcher configures the connection at launch, so a user signs in to a provider, picks a model and starts asking for work. Any other client that speaks authenticated Streamable HTTP MCP has to be given the endpoint and token by hand.
Install and configure
The install path runs through Blender's own extension system: download mcp-for-blender-pro.zip from Releases, open Edit > Preferences > Get Extensions > menu > Install from Disk, enable mcp-for-blender-pro, then open the 3D Viewport sidebar with N under MCP Pro. From that panel a user chooses a provider, adds connection details and selects a tool-calling model. OpenAI is the default; Anthropic Claude, Google Gemini, OpenRouter, OpenCode Zen, Ollama and a custom API are also listed. The same panel can open a separate Studio window.
Two settings matter for anyone running this in a team. Workflow mode is either a single assistant or a chain with one manager, workers and reviewers. Interface language covers English, German, French, Spanish and Chinese.
Limits and open questions
The README names only three of the 27 tools, so what the rest actually do is undocumented outside the code. The quick start is written entirely in PowerShell and the keyword list mentions Windows 3D, but the source never states whether macOS or Linux are supported. Required Blender versions are not given either. Token rotation and revocation are not described, which matters because the token is the only thing gating the endpoint. Whether the manager, worker and reviewer chain multiplies provider usage is not stated. Asset services need optional keys, and the providers behind them are not listed.
What this suggests for AI-assisted builds
Tool servers like this one point at a pattern that is likely to spread: instead of handing an agent a whole machine, a host application publishes a narrow, logged set of operations over a local endpoint. That is a reasonable model for any app you build with AI, where the AI should reach your data through defined actions rather than raw access.
If you are building the app rather than the 3D scene, Draftship takes a plain-language description and produces a hosted app with a built-in database and user sign-in. When you want the app itself to call a model, the guide on adding AI to your app covers how that fits together. Draftship has no connection to this Blender project.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.