VibeInc AI Art Team

An open-source, AI-first Blender toolkit: an "art team" of roles that agents (Claude Code, through MCP) and people (through a Blender panel) can drive. Every action runs in headless Blender too, returns machine-readable JSON, and measures its own output.

Made by VibeInc, an AI-first game studio. The core tool is free and always will be.

GPT concept, then TRELLIS.2 on the local GPU, then Blender cleanup to 12k triangles, 35 cm tall, pivot at the base. Rendered with render_still.

How it works

You, or Claude, ask for something in plain words. The request becomes one of 36 tools, and Blender does the work in the background. Each tool returns what it made, along with its own checks: triangle count, scale, pivot, clearance from walls, and whether anything is cropped. When a check fails, the agent can fix the problem before you ever see it.

The same concept at TRELLIS.2 draft quality (left) and best quality (right), after Blender cleanup.

Why it makes the work easier

• Idea to game-ready prop in about 5 minutes: 45 s of concept art, about 3 min of 3D on a free 16 GB GPU, then seconds of cleanup.
• The boring part is automatic: scale, pivots, UVs, triangle budgets and GLB/FBX exports.
• Blockouts and camera passes in seconds: the same seed gives the same level, and walkthroughs never clip through walls.
• No new subscriptions needed: it uses your ChatGPT plan, your GPU and free Blender. Meshy is an optional paid engine (your own account) for higher-quality meshes, rigging and animation; TRELLIS.2 stays free.
• Agents can run it unattended: results are plain JSON with pass/fail checks, so Claude knows whether each step worked.

What works today (v0.2.1)

Plus: status, list_providers, connect_provider, package_asset, verify_package, about, usage_stats, live_blender, job_status, import_asset, scene_report.

Quick start

Requirements: Blender 4.2 or newer (tested on 5.2.2) and Python 3.10 or newer. The toolkit uses only the standard library, so it has no dependencies.

Install with pip (the Blender-side files come with it):

pip install vibeinc-ai-art-team

Or straight from GitHub: pip install "git+https://github.com/VibeIncStudio/ai-art-team@v0.2.1".

This gives you the vibeart command and the vibeart-mcp server. Machine settings go in ~/.vibeart/local.json (same keys as vibeart.local.example.json), and results go to Documents/VibeInc/art-team-output.

Or work from a checkout:

1. Machine settings. Copy vibeart.local.example.json to vibeart.local.json and set the paths: Blender, your TRELLIS.2 folder, GPU lock files to respect, and your BVH and sound libraries.
2. Check it.

   python -m vibeart.cli status '{}'
   

3. Use it from Claude Code. Copy .mcp.example.json to .mcp.json and fix the path, or add the server by hand:

   claude mcp add vibeinc-art-team -- python <path-to>/ai-art-team/vibeart/mcp_server.py
   

   After a pip install, use claude mcp add vibeinc-art-team -- vibeart-mcp instead.
4. Use it in Blender. Build the extension with python tools/build_addon.py, then in Blender go to Edit > Preferences > Get Extensions > Install from Disk and pick vibeinc_art_team-0.2.1.zip (from the GitHub release, or dist/ after building). The panel is at View3D > Sidebar > VibeInc.

For agents

• Read AGENTS.md. It covers the contract, the rules and the typical flows.
• Result envelope. Every tool returns vibeart.result/1. The status field is one of ok | ok_with_failed_checks | blocked | stub | error. The envelope also carries outputs (file paths), metrics, checks (name, passed, value, expected), warnings and provenance. render_still, cleanup_mesh and image_to_mesh also return the image itself through MCP, so you can look at it.
• Scenes chain. Scene tools take scene (a .blend path). Each step loads it and saves it back.
• Long jobs. concept, image_to_mesh, make_prop, review_asset, improve_asset and make_pack accept background: true, which returns a job_id for job_status.
• Live mode. With the add-on's "Allow live control" preference on, live_blender runs the same actions in the user's open Blender. It listens on localhost only, needs a token, and only runs whitelisted actions.

CLI examples

python -m vibeart.cli greybox '{"scene": "runs/level/level.blend", "preset": "rooms", "seed": 3, "rooms": 5}'

python -m vibeart.cli camera_path '{"scene": "runs/level/level.blend", "mode": "flythrough"}'

python -m vibeart.cli render_still '{"scene": "runs/level/level.blend", "engine": "eevee", "camera": "auto"}'

python -m vibeart.cli make_prop '{"prompt": "a battered field radio", "quality": "draft", "target_height_m": 0.35}'

Art director and asset packs

review_asset sends a render (and, optionally, the concept it came from) to GPT through the Codex app. The job runs in an empty folder with the read-only sandbox, so it can look at the images but cannot change files. The answer must match a fixed JSON schema, and the toolkit checks it again before using it:

• four scores from 1 to 10: silhouette, material_read, matches_brief, game_readiness;
• issues, each with a criterion, a severity (minor, major or blocker) and a suggested fix;
• a one-line summary.

The asset passes when every score is at or above threshold (default 6) and no issue is a blocker. The review is saved as <name>.review-<time>.json (format vibeart.review/1) and recorded in provenance.json with the model, effort, prompt and image hash. One review takes about 15 seconds.

make_pack builds a set of matching props from a theme and a list of items (or a count, and GPT plans the items):

1. Every concept shares one style prefix, so the set looks consistent.
2. Meshes are made one at a time. The step waits up to gpu_wait_s for GPU locks and never runs two TRELLIS jobs at once.
3. Each mesh gets studio lighting and an EEVEE render. The asset .blend itself stays clean.
4. The art director reviews every render. A failing item is regenerated once with a new seed, and the better of the two attempts is kept.
5. A contact sheet PNG shows the whole pack with scores.
6. Everything is packaged as one .vibepack (kind prop) with the merged provenance. TRELLIS output stays commercial_use_cleared: false.

python -m vibeart.cli make_pack '{"theme": "weathered sci-fi field gear", "items": ["field radio", "ammo crate", "medkit"], "background": true}'

Style bible

style_save stores the studio's art direction as styles/<name>.json: a summary, palette, materials, wear level, setting, shape language, things to avoid, and up to three reference images (kept next to the file). Pass its name as style_bible to concept, review_asset, improve_asset or make_pack:

• concepts get the style in the prompt, and the reference images are attached (Codex provider);
• reviews add a fifth score, on_style, and the reviewer sees the references;
• Meshy retextures get the style's materials, palette and wear, which always fit in Meshy's 800-character prompt.

styles/frontier-field-gear.json is an example.

Self-fix loop

improve_asset takes a mesh that failed review and works on its weakest score:

After each fix the mesh is cleaned, rendered and reviewed again, and the best version so far is kept. The loop stops when the asset passes, after max_rounds (default 2), or when the next fix would exceed budget_credits (default 60). Credits are reserved before each request and released only when Meshy refuses it or reports the real charge, so the budget holds even if a request times out. If GPT or Meshy is unavailable, it stops with blocked. In a pack, self_fix: true uses this loop instead of regenerating.

Fixes never change the licence status: anything derived from TRELLIS output stays uncleared, and so does a mesh of unknown origin.

Detail bake (finishing)

bake_detail takes a textured low-poly and a detailed mesh in the same space and bakes the detailed surface into the low-poly's normal map, keeping its colours. replace uses only the baked normals; combine blends them with the low-poly's own normal map. It reports how many pixels carry detail, how many look like wrong hits, and whether the two meshes line up.

With Meshy, image_to_mesh does this by default (finish: true): Meshy remeshes to the triangle budget in the same job (no extra credits in our runs) and also returns its detailed mesh, which is baked on in combine mode. If the bake's own checks fail, the unbaked version is kept and a warning says why. On the field radio the bake took 5 seconds and gave visibly crisper bevels in close-ups.

Gemini (optional, your own key)

Set GEMINI_API_KEY (Google AI Studio, paid tier) to use Google's models:

• review_asset {judges: ["gpt", "gemini"]}: Gemini 4 Argon reviews next to GPT. Scores are averaged, a blocker from either judge fails the asset, and the disagreement is reported. Set "judges": ["gpt", "gemini"] in vibeart.local.json to make it the default. Argon costs about $2/$10 per million input/output tokens until 2026-11-30, then $4/$20.
• concept {provider: "gemini"}: concept art from Nano Banana Pro (gemini-3-pro-image), with style references.
• Fallback: when your ChatGPT usage runs out (or Codex is unavailable), concepts, GPT-drawn views and reviews switch to Gemini automatically and say so in warnings. Set "gemini_fallback": false to turn that off.

The key is read from the environment and only sent to Google; it is never logged or written.

Rigging and animation (Meshy, your own key)

• autorig {source}: Meshy's auto-rigger on a textured humanoid GLB (front toward glTF +Z) or a Meshy task id; 5 credits. You get a rigged GLB and FBX, Meshy's free walk and run clips, and a rig_task_id.
• list_animations {query}: browse Meshy's animation library (free), e.g. "wave" or "walk".
• animate {rig_task_id, actions}: library motions by name or id, 3 credits each, capped by max_credits.

Rigs and library motions are Meshy "Service Assets": use them as part of your character, not on their own. examples/character_demo.py runs the whole chain from a text prompt: GPT views, Meshy multi-view 3D, detail bake, cleanup, art-director review, rig and clips.

LODs

cleanup_mesh, image_to_mesh, improve_asset and make_pack take lods, for example [0.5, 0.25]. That exports <name>_LOD1.glb and <name>_LOD2.glb at half and a quarter of the triangles. Packs ship them by default.

Determinism and self-measurement

• Greybox. A seed or spec always gives the same layout. It reports layout_signature, connectivity, grid snapping and blocks on floor.
• Mesh cleanup. Reports triangles against the budget, UVs in 0–1, UV area (overlap), texel density, pivot at the base centre, transforms applied, height, open edges, export files and a thumbnail. shape_kept fails if triangle reduction shrank the model (it can drop whole regions of some meshes); loose vertices left by the reduction are removed so the size and pivot are measured on real geometry.
• Camera. Reports path length, speed, minimum clearance to geometry and whether the target stays in frame.
• Render. Reports exposure, clipping, subject coverage and render time.
• Photo to PBR. Reports seam ratio before and after, normal validity and height detail.
• Art director. The review format is fixed by a JSON schema and validated locally (one retry if GPT strays). The scores themselves come from a model and can vary between runs.
• Generators. TRELLIS and GPT take seeds or prompts, but their output is not bit-identical between runs. Every run folder keeps provenance.json.

GPU courtesy

image_to_mesh checks the lock files listed in gpu_lock_files and the GPU load first. If another app owns the GPU, it refuses with blocked. The exception is quality: "draft" with allow_busy_gpu: true, which runs a short job of about 3 minutes using about 4 GB. It never kills other processes.

The toolkit also runs only one TRELLIS job at a time on the machine. The TRELLIS worker process itself holds an OS file lock (%TEMP%/vibeart/trellis.lock, shared by the CLI, MCP server and Blender add-on) for exactly as long as it runs, and the OS releases it if the worker dies. make_pack waits for the GPU between items when gpu_wait_s is set.

Money: how VibeInc earns without charging for the tool

1. Referral links. Users connect their own accounts. A connector can carry a VibeInc referral link, which is shown only with this disclosure: "Referral link: VibeInc may earn a referral fee if you subscribe. Your price is the same." Users can switch referral links off. Referral fees never change the ordering or the defaults. See docs/providers.md and docs/affiliate-programmes.md.
2. Marketplace (designed, not launched). .vibepack packages carry licence, provenance, AI disclosure and file hashes. See docs/marketplace.md.
3. Studio funnel. "Made with VibeInc" appears in the panel, in glTF extras.generator and in PNG metadata. A visible render badge is opt-in. The panel also links to the studio and to the opt-in community. Links stay hidden until URLs are set in vibeart/data/branding.json.

Studio and privacy

There is no telemetry unless the user opts in (usage_stats). If they do, the toolkit keeps per-tool counts only, on the user's machine, and has no upload endpoint. Opting out deletes the counts. VibeInc never sells data, never uses the user's machine for background compute, and never trains on users' creations without explicit consent. See docs/privacy.md.

Licences

See LICENSES.md.

• Meshy output (provider: "meshy") belongs to the account holder. On Meshy's free plan it is CC BY 4.0, so credit Meshy. Paid plans give a private licence. Your key stays in MESHY_API_KEY and is never written to disk.
• TRELLIS output is not cleared for commercial use. Its GLB export uses NVIDIA nvdiffrast, which is licensed for non-commercial use only. The toolkit records this in provenance, and package_asset refuses paid listings for it.
• 100STYLE motion needs a credit line (CC BY 4.0).

Tests

cd tests && python -m unittest

• About 100 tests cover every role plus the art director, style bible, self-fix loop, Meshy (generation, finishing, rigging and animation, offline), Gemini, asset packs, LODs, MCP, the add-on install, the live bridge, providers, the marketplace and privacy. The run takes about 6 minutes with headless Blender.
• Live engine tests are opt-in: VIBEART_LIVE_GPT=1 and VIBEART_LIVE_TRELLIS=1. Results of the live runs made during development are in evidence/README.md.

Layout

vibeart/                 host package (stdlib only): roles, tools registry, MCP server, CLI, providers,
                         marketplace, branding, telemetry, jobs, live client
blender_addon/vibeinc_art_team/   Blender extension: ops_core (all bpy actions), headless runner, panel, live bridge
tests/                   unittest suite + Blender fixture builder
tools/                   build_addon.py, try_action.py, when_gpu_free.py, showcase/ (renders + composes the demo videos)
examples/                character_demo.py (concept to rigged, animated character)
docs/                    providers, affiliate research, marketplace design, privacy
evidence/                live run outputs (GPT concept, TRELLIS mesh, renders)

Inspiration

Scenario's Blender plugin (announced by @emmanuel_2m on X, 14 Oct 2026) showed eight AI roles exposed through MCP. This project has no Scenario code and does not use Scenario's service. It follows the same idea, but with engines you run yourself or connect with your own accounts.