blender-ai-copilot-lab

An AI lab for Blender where manager–worker–reviewer agent teams run multi-step scene briefs as supervised experiments—a live task board, review findings, one correction attempt per blocked task, named undo, and a built-in MCP server.

[]

Download

Set the brief and watch the board: the manager splits it, workers build, reviewers return findings—and one approved correction later, the scene passes.

Why

Big briefs die in a single chat thread—no structure, no checks, no record of what actually failed. blender-ai-copilot-lab turns the brief into a supervised experiment: a manager agent breaks it into assignments with acceptance criteria, worker agents execute each one as ordered scene operations, and reviewer agents approve the result or return specific findings. You watch the whole run on the task board, correct once when a task blocks, and keep a clean, undoable scene at the end of every experiment.

Without / With

Without	With
One endless chat thread that loses the plan	A manager that splits the brief into assignments and acceptance criteria
No idea what the agent did or why	A task board showing assignments, attempts and review findings
Failed steps silently skipped	One correction attempt per blocked task, then an honest stop
Edits landing out of order	Tasks run sequentially so scene edits stay ordered
Review by eyeballing only	Reviewer agents that approve or return specific findings
Doing it all solo anyway	A team chain of named agents, each with its own model

Quick Start

Every command below is self-contained: open PowerShell, copy it, paste it, press Enter.

Download mcp-for-blender-pro.zip from Releases.
In Blender, open Edit → Preferences → Get Extensions → menu → Install from Disk.
Select the ZIP, enable mcp-for-blender-pro, then open 3D Viewport → N → MCP Pro.
Choose a provider, add your connection details and select a model.
Send a message in Blender or click Open Studio Window.

Features

Break briefs into assignments—a manager agent defines tasks and acceptance criteria before anything moves.
Build with worker agents that execute each assignment through 27 supervised scene tools.
Review every result—reviewer agents approve work or return specific findings.
Watch the run live—assignments, attempts and findings stay visible on the task board.
Correct once—a blocked task gets exactly one correction attempt, then the chain stops honestly.
Configure each agent—names, responsibilities, instructions, providers and models per agent.
Keep every edit reversible—named undo steps and per-operation approvals across the whole run.

How it works

blender-ai-copilot-lab runs inside Blender as an extension: your brief becomes a sequence on the task board—a manager agent splits it into assignments with acceptance criteria, worker agents execute them through 27 supervised scene tools, and reviewer agents approve each result or return findings. A blocked task gets one correction attempt; a second failed review stops the chain after the current step, and every mutation leaves a named undo step.

Usage

A staged environment brief. "Build a product-display scene: pedestal and camera first, three lights second, matte blue material last." The manager sequences the stages, workers deliver each one, reviewers sign off in order.

A review loop you can read. A reviewer returns "pedestal material too glossy"—the worker gets one correction attempt, and the finding stays on the board for you to inspect.

Mixed models per role. Give workers a local Ollama model for the heavy lifting and the reviewer a cloud model for judgment; each agent keeps its own connection.

Configuration

Setting	Default	What it does
Workflow mode	Single assistant	Single assistant for plain chat, or a Team chain for supervised experiments
Team chain	One manager, workers, reviewers	Exactly one enabled manager, one or more workers, one or more reviewers
Agent fields	Per agent	Name, responsibility, instructions, provider, model, manager and enabled switch
Model inheritance	Workspace setting	Blank provider and model fields fall back to the workspace connection
Provider	OpenAI	OpenAI, Anthropic Claude, Google Gemini, OpenRouter, OpenCode Zen, Ollama or Custom API
MCP endpoint	`http://127.0.0.1:9876/mcp`	Local Streamable HTTP endpoint for external MCP clients; enable in extension preferences
Access token	Generated when enabled	Bearer token for the MCP endpoint; keep it in the client's secret store
Interface language	English	English, German, French, Spanish or Chinese for the main interface
Asset services	Optional keys	Service keys for asset providers that require them
Custom endpoint	One shared	The single saved custom endpoint is shared by custom-protocol agents

Troubleshooting / FAQ

How many managers can I enable? Exactly one—the chain refuses to run with two enabled managers, so responsibilities stay clear.

What happens on a second failed review? The chain stops after the current step; nothing fails silently, and the board shows both findings.

Can agents use different models? Yes—each agent has its own provider and model; blank fields inherit the workspace setting.

Do tasks run in parallel? No—tasks run sequentially so scene edits stay ordered and reviewable.

Can I work without a team? Yes—Single assistant keeps the normal chat workflow for quick edits.

Contributing

Spotted a chain behaving badly? Open an issue with the brief, the board state and what you expected, or send a pull request. If blender-ai-copilot-lab saves you an afternoon of redo loops, a star helps other artists find it.

Keywords

blender-ai-copilot-lab · blender · blender 3d · blender addon · blender plugin · blender extension · blender scripting · blender automation · blender mcp · blender ai · blender assistant · blender copilot · blender chatgpt · blender llm · blender agents · bpy · 3d modeling · 3d automation · mcp · mcp server · model context protocol · ai for blender · agent teams · multi agent · task board · manager worker reviewer · agent review · acceptance criteria · scene tools · scene generation · scene inspection · ai modeling · ai materials · viewport screenshot · studio workspace · product visualization · geometry nodes · 3d workflow · 3d rendering · 3d scene builder · windows 3d · ai 3d · ollama · openai · anthropic claude · google gemini · opencode zen · model catalog · connection profiles · custom api · poly haven · sketchfab · asset search · blender workflow · chat in blender · ai chat blender

License

Released under the MIT license.