AI Chat Panel & Providers

Beyond acting as an MCP server for external clients, BlenderMCP Pro includes a complete agentic chat interface built directly into the N-panel. It can call every one of the 75 tools itself, so you can drive Blender from natural language without ever opening a separate application.

Choosing a Provider

Open Edit → Preferences → Add-ons → BlenderMCP Pro → Preferences and pick an AI Provider. Five are supported:

Claude (Anthropic) — the most reliable tool use of the five. Requires an Anthropic API key from console.anthropic.com.
Gemini Flash (Google) — a generous free tier (1M tokens/day). Requires a free key from aistudio.google.com. Defaults to gemini-2.5-flash; the model field is editable if you want to try gemini-2.0-flash or another Gemini model.
Groq — free tier, very fast inference on Llama 3.1 70B. Requires a free key from console.groq.com.
GPT-4o-mini (OpenAI) — reliable tool use at low cost. Requires a key from platform.openai.com.
Local LLM (Ollama) — fully free and offline, with no data leaving your machine. Tool-calling reliability depends heavily on model size; see below.

Using a Local Model (Ollama)

Selecting Ollama shows a live status indicator ("Running" / "Not detected") and a model field with a refresh button that lists models already pulled locally. A Pull Model button is available to download a new one without leaving Blender. Recommended models, in order of capability:

qwen2.5:0.5b — very fast, weak at tool use (~400MB)
qwen2.5:3b — best balance for CPU-only machines (~1.9GB), the default
qwen2.5:7b — noticeably stronger tool use (~4.4GB), GPU recommended
llama3.1:8b — most capable of the four (~4.7GB), needs a GPU to be responsive

Because small local models are unreliable at structured tool calls, the chat engine caps Ollama at 3 tool-use loops per message (versus 20 for the cloud providers) and gives a clearer timeout message if a local model stalls.

Using the Chat Tab

Open the Chat tab in the MCP Pro N-panel. Type a request in natural language — for example "add a three-point lighting rig" or "decimate the selected mesh to under 5000 faces and export it as FBX" — and the AI will plan and execute the necessary tool calls against your scene automatically, showing each call and its result inline as it works.

Viewport Screenshot Input

The chat panel can attach a screenshot of the current 3D viewport to your message, so the AI can see what you're looking at rather than working blind from scene data alone. The image is automatically downscaled to a maximum width of 1280px and compressed to JPEG before being sent, keeping requests fast and token-efficient. This requires a provider with vision support (Claude, Gemini, and GPT-4o-mini all qualify).

Voice Input

Click the microphone icon to dictate a chat message instead of typing it. See Voice Input for setup requirements.

Token Optimizer

To keep long conversations fast and affordable, the chat engine automatically:

Prunes the tool schema sent to the AI down to the categories relevant to your current message, rather than always sending the full 75-tool schema.
Compresses scene context to only what's relevant to the conversation.
Compresses older chat history after 6 full turns, summarizing rather than resending full transcripts.
De-duplicates repeated tool-call IDs and caches the tool schema between requests where possible.

None of this is configurable — it runs automatically in the background on every request.

Sessions

Each conversation is a session, automatically titled from your first message. Use New Session in the Chat tab to start fresh at any point; previous sessions aren't lost. See Chat History & Sessions for browsing, renaming, favoriting, and deleting past sessions.