Agent Mode

Agent Mode lets you control Quadify Ultra in plain English. Type an instruction like "Full remesh this boulder scan" or "Fix n-gons and quadify everything", and the Agent plans and executes the right operations automatically. Switch to the Agent tab using the mode selector at the top of the Quadify Ultra panel.

Setup

1. Enable Agent Mode using the toggle in the panel.
2. Select a provider: Claude (Anthropic) or OpenAI.
3. Paste your API key in the API Key field. The key is stored locally in Blender preferences and never transmitted except to the chosen provider.
4. Choose a model quality: Fast (~$0.001/run, Claude Haiku — recommended for most uses) or Quality (~$0.020/run, Claude Sonnet — for complex multi-step plans).

Scene analysis uses a single batched call for all selected objects regardless of count, keeping costs predictable.

Writing Instructions

Instructions are plain English. A few patterns that work well:

• "Quadify the entire mesh" — processes the active object
• "Full remesh this boulder scan" — routes to QuadWild or Voxel Remesh based on mesh type
• "Use QuadWild on the character" — explicit algorithm selection
• "Fix n-gons and quadify everything" — multi-step plan: fix n-gons then convert
• "entire" or "all" — processes every visible object automatically

The Agent remembers the last two exchanges as conversation context, so follow-up instructions like "now reduce the poly count by half" work without repeating the mesh name.

Preview Before Run

Enable Preview before run to see the plan before executing it. The plan card shows:

• Number of steps and a plain-English summary
• Per-step confidence %, whether the step is destructive, and the operation name
• Any warnings (shape keys detected, mirror modifier, high face count)

Click Execute Plan to run it or Discard to cancel. Preview mode is recommended for unfamiliar mesh types or complex multi-step instructions.

Prompt History

Every instruction that runs is auto-saved to prompt history (last 20 entries). From the history panel you can:

• Load any past prompt back into the instruction field with one click
• Star prompts as favourites (shown with a gold icon)
• Filter the list to favourites only
• Delete individual entries

Use the + and ★ buttons next to the instruction field to save or favourite the current instruction before running.

Analyse Entire Scene

The Analyse Entire Scene button sends anonymous mesh statistics for all visible objects and returns a scene-wide summary: which objects need work, which are already clean, and a suggested processing order. No instructions are executed — this is a read-only analysis call.

Conversation History

The Agent saves the last 40 exchanges to logs/quadify_ultra_history.json inside the addon directory. History loads automatically on Blender open, so multi-session workflows continue naturally. The trash icon next to "Continuing conversation" clears both in-memory and file history. Use Clear History (↺ button in the panel header) to start a fresh conversation without prior context.

Advanced Settings

• Model Override — enter any model string (e.g. claude-opus-4-6, gpt-4o) to pin to a specific version. Leave empty to use the default for the selected quality.
• Enable MCP (Experimental) — sends the anthropic-beta: mcp-client-2025-01-01 header with every Claude request. Only useful if you have a local Blender MCP server running. Claude provider only. Off by default.

Privacy

Only instruction text and anonymous mesh statistics (face counts, topology ratios, algorithm outcomes) are transmitted. No mesh geometry, vertex positions, filenames, object names, or personal data is ever sent.

If the Agent Gets Stuck

If a run appears to hang, click Agent stuck? Click to reset in the panel to clear the running state without restarting Blender. Then retry the instruction, optionally with a simpler phrasing.