game-asset-mcp
A local-first MCP server for producing, inspecting, repairing and shipping game assets, with paid provider calls gated behind spend limits and many fully free local tools.
Plan and preview prompts for free —
preview_asset_promptshows the exact reconstruction-oriented prompt/negative prompt a spec would send, no credits spent.Generate reference images (spends image credits) —
generate_asset_referencemakes reconstruction-framed candidate images from a structured spec;create_game_propis a one-shot brief-to-candidates shortcut;generate_reference_variationsexplores one named design axis on a new child job.Pick and reconstruct 3D (spends 3D credits) —
select_referencerecords the chosen candidate locally, thencreate_3d_assetreconstructs from that job, a local image, an https URL, or text only.Extend existing models (spends 3D credits) —
texture_existing_assetadds PBR materials to a mesh without touching geometry;rig_assetbuilds skeletons/skin weights;animate_assetretargets preset animations onto rigged assets;retopologize_assetrebuilds topology, quads by default.Generate audio (spends audio credits) —
generate_sound_effectmakes short SFX, impacts, UI blips or seamless loops from text.Manage jobs and download results —
get_asset_jobpolls status (read-only, no credits),list_asset_jobslists local jobs, anddownload_assetpulls the model, PBR variant, preview and reference into a workspace withasset.jsonprovenance, extracting embedded textures.Inspect and judge assets locally and free —
inspect_assetreports meshes, triangles, materials, texture sizes, bounding box and missing channels;validate_game_assetreturns a pass/fail game-readiness verdict against an overridable shipping policy.Repair and process meshes locally and free —
normalize_meshgenerates missing UVs, welds, cleans geometry and optionally decimates with a before/after receipt;batch_prepare_meshesvalidates, normalizes and re-validates a list in one call;extract_pbr_triosplits a material into separate albedo/normal/roughness planes with a channel receipt.Track spending —
get_spend_reportshows estimated provider costs by tool and remaining headroom under the configured spend limit, with no network call.
Allows using a local Blender installation to repair and normalize meshes for asset preparation, including the repair half of normalize_mesh and batch_prepare_meshes.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@game-asset-mcp Generate a 3D asset from my reference image with PBR textures."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Game Development Studio
Local-first tools for producing, packaging, vendoring, capturing, debugging,
and optimizing game assets and renders. The stable automation interface is the
game-dev command-line program and its JSON/JSONL protocol; five focused skills
and a native macOS companion build on that same inspectable boundary.
The project is designed for human developers and local coding agents that need more than a screenshot: immutable run bundles can combine render attachments, scene controls, logs, counters, timing samples, and provenance so a visual diagnosis has inspectable evidence behind it.
What it does
Drives Tripo 3D and Leonardo image/audio jobs through explicit, per-invocation spend authorization.
Inspects and validates GLB assets, normalizes meshes through Blender, and creates optional USDZ previews.
Builds content-addressed asset packages with hashes, receipts, provenance, licenses, validation results, and a rebuildable catalog.
Plans project vendoring before it writes and blocks unknown licenses, invalid packages, symlink escapes, and destination collisions by default.
Runs declarative, project-owned capture scenarios with independent authorizations for execution, GPU use, and hardware-performance collection.
Seals render outputs, semantic attachments, telemetry, logs, and metrics into verifiable run bundles.
Computes deterministic raster statistics, heatmaps, attachment-aware comparisons, metric summaries, and bounded optimization goals.
Ships a router and four focused Codex/ChatGPT skills without automatically installing anything into a user profile.
Related MCP server: Context3D MCP Server
Requirements
Node.js 22.5 or newer
macOS, Linux, or Windows for the CLI
Blender only for normalization or Blender-backed preview workflows
Tripo or Leonardo credentials only for the corresponding paid provider operations
A project adapter only when executing game-specific capture scenarios
The native companion requires macOS 26 and a Swift 6.2 toolchain to build from source. It remains optional: the CLI and skills do not require the app.
Repository layout
Path | Owns |
| Cross-platform |
| Router plus four focused Codex/ChatGPT skills |
| Declarative game-specific capture adapters |
| Native macOS 26 SwiftUI companion |
| Protocol, architecture, package, adapter, and app contracts |
| Reproducible public skills and binary-app repository templates |
| Byte-provenanced product art, screenshots, and publication copy |
See Repository layout for what each distribution ships and which outputs are generated rather than source-controlled.
Product tour

Product composition using the shipped skill names and metadata.

Marketing composition based on actual v1.0.0 CLI output, shortened for display. It illustrates the command contract; it is not a current-run claim.

The third image is explicitly a synthetic validation fixture. It demonstrates the diagnostic layout and evidence language; it is not a target-game capture, hardware GPU result, or performance claim.
Native macOS app
The repository also contains a native SwiftUI companion with four first-class
workspaces: Production, Library & Vendoring, Visual Debugging, and Performance.
It presents the existing local game-dev protocol through a
NavigationSplitView, system search and toolbar controls, a result inspector,
keyboard commands, a dedicated Settings scene, and explicit empty, loading, and
error states. Its complete bundle carries a closed, exact-rostered build of the
same CLI plus a direct Node runtime; it has no publisher-hosted backend.
Provider credentials are entered in masked fields, stored in the macOS Keychain, and shown only as configured or not configured. Paid provider calls, package construction, project vendoring, scenario execution, GPU capture, and hardware-performance collection use one-shot approval sheets. Vendoring and scenario execution are dry-run-first, and changing their planned inputs invalidates the current approval path. Sensitive operations bind that approval to the full local runtime-tree digest and execute a separately verified private snapshot, rather than trusting an arbitrary global CLI or PATH-selected Node.
From the repository root:
./script/build_and_run.sh --test
./script/build_and_run.sh --build-only
./script/build_and_run.sh # build and launch
./script/build_and_run.sh --verify # process-level launch check onlyThe helper produces
apps/macos/GameDevelopmentStudio/dist/GameDevelopmentStudio.app with an ad-hoc
signature for local development. It is not a Developer ID-signed, notarized, or
Mac App Store build. The --verify mode finding a live process after two
seconds does not prove window readiness, pixel correctness, workflow behavior,
accessibility, or human acceptance.

This is a full-resolution capture of one reviewed native macOS 26 runtime state:
the Visual Debugging workspace in the default Dark appearance after a successful
local doctor check. Its byte-level provenance is recorded separately. It does
not prove a target-game GPU capture, pixel correctness across the app, light-mode
behavior, accessibility, performance, signing identity, notarization, or human
acceptance.
See Native macOS app for the architecture, approval and Keychain boundaries, commands, local-bundle status, and exact evidence limits.
Install
The CLI source is public. Release 1.5.1 provides a compiled Node.js package
(theisegoria-game-development-studio-1.5.1.tgz) and SHA-256 manifest through
GitHub Releases.
It requires a separately installed Node.js 22.5+ runtime; it is not a standalone
Windows EXE/MSI. The skills plugin ZIP does not install game-dev.
The public npm registry package is not published. Use the GitHub release tarball or source build. Follow the canonical install guide for checksum verification, POSIX/PowerShell installation, PATH, manual update and rollback. No native macOS app is required.
From source:
git clone https://github.com/theisegoria/game-development-studio.git
cd game-development-studio
npm ci
npm run build
node dist/cli.js --helpNo provider call is made during installation, build, test, doctor, or
capabilities.
Quick start
New here? Your first capture goes from an empty directory to a sealed synthetic capture your AI can verify and compare. The built-in generic sample requires no compiler, provider credentials or GPU. Release 1.4.0 runs the sample through an explicit Node interpreter on Windows, macOS and Linux. Engine compilation is covered separately in the advanced C probe tutorial.
Choose a workspace and inspect the local environment:
game-dev capabilities --output-dir ./asset-workspace --json
game-dev doctor --output-dir ./asset-workspace --json
game-dev credentials status --output-dir ./asset-workspace --jsonInspect an existing GLB without invoking Blender or a provider:
game-dev asset inspect ./model.glb --output-dir ./asset-workspace --json
game-dev asset validate ./model.glb --output-dir ./asset-workspace --jsonBuild and verify a canonical package. package build writes a new package even
though it has no --confirm flag, so review the resolved source, output
workspace, name, version, license, and collision behavior before authorizing
that exact invocation:
game-dev package build ./model.glb \
--name "Signal Beacon" \
--version 1.0.0 \
--license CC0-1.0 \
--output-dir ./asset-workspace \
--json
game-dev catalog list --output-dir ./asset-workspace --json
game-dev package verify PACKAGE_ID --output-dir ./asset-workspace --jsonProject admission is dry-run-first:
game-dev vendor admit PACKAGE_ID --project /path/to/game --json
game-dev vendor admit PACKAGE_ID --project /path/to/game --confirm --jsonThe second command is a new invocation with explicit write authorization. A plan does not become standing permission.
Paid provider jobs
Provider routes require game-dev 1.0.2 or newer. The account holder must
configure the provider credential outside ChatGPT, Codex, and the plugin
conversation, using a local credential mechanism they control. The CLI reads a
preconfigured environment value lazily and never accepts a key as a command
argument; the native macOS app can use Keychain. Never paste a key into a
conversation, request file, source file, log, screenshot, or support report.
Check presence without revealing the value:
game-dev credentials status --output-dir ./asset-workspace --jsonEvery paid invocation needs both approval and a finite estimated spend ceiling:
game-dev provider tripo generate \
--request ./requests/prop.json \
--approve-spend \
--spend-limit-cents 100 \
--output-dir ./asset-workspace \
--jsonlThese are optional user-operated requests made directly from the local CLI to the selected provider using the user's own account and API access. The publisher runs no proxy, shared account, credential service, resale layer, or provider job queue. Supported routes are:
Tripo:
generate,retexture,rig,retarget, andretopologizeLeonardo:
image-generateandsound-generate
The ceiling is a refusal guard based on estimated prices, not a provider invoice. Provider submission creates a durable local job before polling so an interruption does not erase what was requested or what may have been charged. Resuming a job requires fresh authorization.
Capture harness
An adapter is a declarative .game-dev/adapter.json owned by the target game.
It names scenarios, commands, parameters, declared capabilities, and the run
output contract. Installing a template and inspecting an adapter do not execute
the game.
game-dev adapter templates --json
game-dev adapter install genome-game --project /path/to/game --json
game-dev adapter install genome-game --project /path/to/game --confirm --json
game-dev scenario list --project /path/to/game --json
game-dev scenario plan trident-bay-contract --project /path/to/game --jsonExecution requires --confirm. A scenario that declares GPU or
hardware-performance capability additionally requires --allow-gpu or
--allow-performance for that invocation:
game-dev scenario run trident-bay-windowless-metal \
--project /path/to/game \
--request ./capture-parameters.json \
--confirm \
--allow-gpu \
--jsonlOnce the adapter has produced its staging output, the harness validates the closed artifact roster and hashes it into a sealed run bundle. Typical attachments include color, depth, normals, object IDs, material IDs, motion, and overdraw, but the adapter declares the exact set.
game-dev capture verify RUN_ID --json
game-dev visual analyze RUN_ID --json
game-dev visual compare BASELINE_RUN CANDIDATE_RUN \
--threshold 0 \
--output ./comparison \
--jsonl
game-dev performance compare BASELINE_RUN CANDIDATE_RUN --stat median --jsonSee Capture adapters for the manifest and output contract.
Skill suite
The distribution contains five self-contained skills:
game-development-studio: router and shared operating contractgame-asset-production: provider, inspection, normalization, and packaginggame-asset-vendoring: catalog, integrity, migration, and project admissiongame-visual-debugging: adapters, captures, telemetry, and raster evidencegame-performance-optimization: metrics and bounded optimization goals
List the exact packaged bytes:
game-dev skill list --jsonInstallation is dry-run-first and refuses symlinked targets or drifted existing copies:
game-dev skill install all --target /path/to/codex/skills --json
game-dev skill install all --target /path/to/codex/skills --confirm --jsonNothing in this repository installs into ~/.codex, Genome, or another game
project without an explicit confirmed command. The standalone plugin source is
published separately at
theisegoria/game-development-studio-skills.
Machine-readable protocol
--jsonemits exactly onegame_dev.result.v1object to stdout.--jsonlemits orderedgame_dev.event.v1records followed by one terminal result event.Logs and diagnostics go to stderr.
Secrets are redacted from structured errors, receipts, jobs, and URLs.
Persisted artifacts use atomic replacement and closed hash rosters where the format promises immutability.
Read CLI protocol before integrating the command with an agent or GUI.
MCP server
The same local operations are also served over MCP on stdio, for AI clients that have no shell — Claude Desktop, Codex, Gemini, and any other MCP client. It is a second transport over one registry, not a second implementation: the same tools, contracts, spend rules and evidence ceilings, with a release check that fails if the two ever advertise different tool sets.
game-dev mcp config --client claude-desktop
game-dev mcp config --client codex --spend-limit-cents 500Generate the configuration rather than writing it by hand: ASSET_OUTPUT_DIR
must be absolute, because a relative path resolves against a working directory
the client chose, and several clients spawn servers from /.
Paid tools are disabled unless the configuration a human wrote supplies a spend ceiling, and each charge then requires approving an elicitation prompt. No tool takes an approval argument, because over MCP the model writes every argument. See MCP server.
Workspace layout
By default, ASSET_OUTPUT_DIR is ./assets/generated; durable state is
under $ASSET_OUTPUT_DIR/.game-dev.
assets/generated/
├── .jobs/ legacy-compatible asset job records
└── .game-dev/
├── jobs/ durable provider operations
├── packages/ canonical game asset packages
├── runs/ sealed capture bundles
└── catalog.sqlite3 rebuildable derived indexSet GAME_DEV_DATA_ROOT to place durable state elsewhere. See
Asset packages and Architecture.
Trust and evidence
Game Development Studio deliberately distinguishes evidence classes:
Source inspection proves only what is present in source.
Typecheck, unit, and contract tests prove those checks in the tested environment.
Blender-backed tests prove the exercised headless Blender workflows.
A valid capture bundle proves its declared files, hashes, controls, and adapter-reported telemetry.
Only a real admitted run can support hardware GPU or performance claims.
None of those automatically proves pixel correctness, causality, signing, notarization, or human visual approval.
Provider tests use local HTTPS fixtures and do not spend credits. Live-provider acceptance, target-hardware captures, and human review remain separate gates.
Development
npm ci
npm run typecheck
npm run lint
npm test
npm run verify
npm pack --dry-run --jsonThe full local suite may need permission to bind a loopback HTTPS fixture. Real Blender/GPU tests require their explicit opt-in switches and are independently enforced in CI. Optional Basis and CoACD CPU verification uses dedicated CI lanes; default tests mock those native backends.
See Contributing, Security, Privacy, and Terms.
License
MIT © 2026 Benjamin Michael Haire. Third-party providers, generated content, source assets, and vendored assets remain subject to their own terms and licenses.
Catch visual regressions in pull requests
Add the visual regression gate to your engine's repository: every pull request is captured against its base commit on the same runner, compared beyond the measured noise floor, and summarised with heatmaps.
- uses: theisegoria/game-development-studio/action@v1.5.1
with:
scenario: captureStandalone production workflows (1.3)
All new operations are discoverable through game-dev capabilities --json and
usable through game-dev tool call NAME --input JSON --json or the same MCP name.
Mutations and recorded human decisions require --confirm per CLI invocation;
MCP requires human elicitation. Recipe provider steps additionally require fresh
spend approval and a ceiling. An approval fingerprint is an input binding, never
standing permission.
Accounting and recovery: concurrent reservations, corrupt/stale job diagnostics, provider cost/outcome/quality history.
Blender reliability: typed receipts, checkpointed batches, explicit test opt-in, process intent and bounded resource use.
Asset review: local CPU candidate views, reviewed packages, named baselines, scenario matrices and expected-change history.
Workspace retention and updates: measured storage, protected evidence, reversible quarantine, folder exports, separately confirmed irreversible quarantine purge, and verified same-distribution GitHub upgrade/rollback plans for CLI, skills and the existing Anvil ARM64 archive.
Production recipes: checkpointed workflow graphs, approved sample families, standalone platform variants, LODs and collision proxies.
CPU texture compression: ETC1S/UASTC KTX2 GLBs, material-aware color handling and per-mip CPU verification before packaging.
CPU convex decomposition: separate validated convex parts, bounded native processing and sampled approximation evidence.
Basis and CoACD require explicitly configured optional dependencies; they are not automatically installed or bundled into the native CLI runtime. Missing dependencies block their operations without blocking ordinary inspection. Purge permanently removes reviewed quarantined files and requires stopped writers and fresh explicit consent; reported logical bytes are not a measurement of filesystem space reclaimed.
These workflows do not integrate a game engine. CPU previews are static inspection,
not final rendering or quality certification. Unsupported conversions remain explicit.
Local fast tests use npm test; real Blender and GPU lanes require the opt-in
switches documented above and run in CI for release validation.
Visual and performance completion upgrade
The native companion provides run discovery, typed scenario parameters, attachment viewing, synchronized visual comparison, sample/distribution charts, and reviewable external-agent optimization sessions. Analysis v2 distinguishes raw samples from supplied aggregates and reports missing or incompatible data.
Create a disposable generic project with
game-dev adapter sample --project NEW_DIRECTORY --confirm --json.
Read the optimization session guide for the
complete snapshot, build/test, capture, evaluation, and patch-review workflow.
Synthetic samples demonstrate the workflow; they are not game-hardware results.
Available Tools
20 toolsanimate_assetRetarget an animation onto a rigged assetA
SPENDS 3D CREDITS. Applies a preset animation to an asset that has ALREADY been rigged with rig_asset. Each animation costs separately. ASYNCHRONOUS: poll get_asset_job, then download_asset. Refuses when the source job was not a rig, because retargeting an unrigged model produces nothing while still being billed. NOTE: the provider task name is taken from published docs and is UNVERIFIED live.
Inputs: animation (required), assetJobId (required), outFormat.
batch_prepare_meshesNormalize and validate many meshes in one callA
FREE and fully local: no network call, no credits. Runs the preparation pipeline over a list of meshes — validate, normalize the ones that fail, validate again — and returns a per-item verdict plus a summary. Meshes that already pass are left untouched rather than rewritten. One failing item never stops the run; its error is reported and the batch continues. Normalization needs a local Blender install; without one this still validates and simply reports what would need repairing. Note that a FAILED item can still have written a file: when normalization succeeds but the result does not clear the policy, the mesh is kept for inspection and named in normalizedPath with outputKept set. Use outputsWritten, not prepared, to predict how many files are in the output directory — prepared is a verdict, not a file count.
Inputs: modelPaths (required), normalize, outputDir, requireUVs, +4 more.
create_3d_assetReconstruct a 3D assetA
SPENDS 3D CREDITS. Creates a textured 3D model. ASYNCHRONOUS — returns immediately with a job id; poll get_asset_job until status is "ready", then call download_asset. Provide exactly ONE source: assetJobId (uses that job's selected reference image), imagePath (a local png/jpg/webp), imageUrl (https), or textPrompt (skips images entirely). Model URLs from the provider EXPIRE, so download promptly once ready.
Inputs: pbr, quad, seed, spec, +7 more.
create_game_propStart a game prop from a briefA
SPENDS IMAGE CREDITS ONLY. Convenience entry point: takes a loose brief, builds a reconstruction-grade prompt, and generates N candidate reference images. It deliberately STOPS THERE and does NOT spend 3D credits — look at the returned candidates, call select_reference with your choice, then create_3d_asset. This exists so art direction stays with you rather than being guessed at.
Inputs: name (required), description (required), category, numImages, +5 more.
download_assetDownload a finished asset into the workspaceA
FREE: no credits. Downloads a completed asset job into a local workspace folder — the model, any PBR variant, the provider preview, the source reference image — and EXTRACTS embedded textures into textures/ as separate files. Also writes asset.json with complete provenance. Call this as soon as get_asset_job reports "ready": provider URLs expire. Never overwrites an existing workspace; a repeat download gets a new suffixed folder.
Inputs: assetJobId (required), destination, extractTextures.
extract_pbr_trioSplit a material into separate PBR texture planesA
FREE and fully local: no network call, no credits. Splits a glTF material into independent albedo, normal and roughness images (plus metallic and occlusion when present), de-packing the glTF metallicRoughness texture — roughness is its GREEN channel, metallic its BLUE. Resamples to an exact resolution, averaging colour in linear light and data channels directly, and writes a receipt naming the source texture and channel behind every plane. Use after download_asset, or on any local glTF/GLB. Planes a material declares only as a scalar factor are emitted as flat images and marked source="factor" so a constant is never mistaken for measured data.
Inputs: modelPath, assetJobId, resolution, destination, +1 more.
generate_asset_referenceGenerate reference images for a game assetA
SPENDS IMAGE CREDITS. Generates candidate reference images for a game asset, framed for 3D reconstruction rather than as concept art (isolated object, whole silhouette, neutral background, even lighting). Returns candidate image URLs for YOU to look at. It does NOT create a 3D model — after inspecting the candidates, call select_reference then create_3d_asset. Asynchronous: waits up to waitSeconds, then returns a job id to poll.
Inputs: spec (required), seed, width, height, +3 more.
generate_reference_variationsExplore variations of an existing referenceA
SPENDS IMAGE CREDITS. Generates variations of an existing asset job along ONE named axis (silhouette, material_treatment, industrial_detailing, wear, proportions, functional_components) while holding the object identity fixed — a variation should still be the same object, not a different one. Creates a NEW child job linked to the parent.
Inputs: axis (required), assetJobId (required), seed, width, +4 more.
generate_sound_effectGenerate a game sound effectA
SPENDS AUDIO CREDITS. Generates a short sound effect from a text description — weapon reports, impacts, footsteps, UI blips, or a seamless ambience loop. Set loop:true for anything that must tile without a seam. Each variation in quantity costs separately. Polls and downloads inline, returning local file paths; the wait is bounded, so a slow provider returns a job id to poll with get_asset_job rather than hanging. NOTE: the provider documents this request contract but not its response shape, so this path is UNVERIFIED against the live API — the first real call may need a client fix.
Inputs: name (required), prompt (required), loop, quantity, +4 more.
get_asset_jobGet asset job statusARead-only
Check an asset job, refreshing it from the provider if it is still running. READ-ONLY: makes a network call to the provider but never spends credits. Call this after create_3d_asset or texture_existing_asset until status is "ready", then call download_asset. Statuses: queued, generating_reference, reference_ready, generating_3d, processing, ready, failed, cancelled.
Inputs: assetJobId (required), detail.
get_spend_reportReport what this session has spentARead-onlyIdempotent
FREE and fully local: no network call, no credits. Reports the estimated provider spend for this workspace, broken down by tool, plus the remaining headroom under ASSET_SPEND_LIMIT_CENTS if one is set. Figures are normalised to US cents because providers bill in different units. Costs marked "estimated" are pessimistic placeholders for providers that do not publish a per-call rate — treat them as a guard, not an invoice.
Inputs: no named parameters.
inspect_assetInspect a downloaded 3D assetARead-onlyIdempotent
FREE and fully local: no network call, no credits. Reports the technical properties of a downloaded glTF/GLB — mesh and triangle counts, materials, texture resolutions, bounding box in metres, and whether UVs, normals, tangents and PBR channels are present — plus warnings for anything that will cause downstream problems. Pass assetJobId to inspect a downloaded job, or modelPath to inspect any local file.
Inputs: modelPath, assetJobId.
list_asset_jobsList asset jobsARead-only
List local asset jobs, newest first. READ-ONLY and fully local: no network call, no credits. Does not refresh provider state — use get_asset_job for that.
Inputs: limit, status.
normalize_meshRepair a mesh so it can be textured and shippedA
FREE and fully local: no network call, no credits. Requires a local Blender install (optional dependency; the tool refuses with instructions if absent). Repairs a mesh for downstream use: generates UV coordinates for objects that have NONE — the single most common reason a mesh cannot be textured — welds coincident vertices, dissolves degenerate triangles, gives every material a stable non-empty name, forces opaque blending, and optionally decimates to a triangle budget. Existing UV layouts are never overwritten. Exports GLB and returns a receipt with before/after counts so the change is measurable.
Inputs: modelPath, overwrite, assetJobId, outputPath, +8 more.
preview_asset_promptPreview the prompt that would be generatedARead-onlyIdempotent
FREE and fully local: no network call, no credits. Shows the exact reconstruction-oriented prompt and negative prompt that generate_asset_reference would send for a given spec. Use this to sanity-check art direction before spending image credits.
Inputs: spec (required).
retopologize_assetRebuild an asset's topology, optionally as quadsA
SPENDS 3D CREDITS. Rebuilds the topology of an existing generated asset. Quad output is the default because quads survive downstream mesh qualification and editing far better than the triangle soup generators typically emit. ASYNCHRONOUS: poll get_asset_job, then download_asset and validate_game_asset. NOTE: the provider task name is taken from published docs and is UNVERIFIED live.
Inputs: assetJobId (required), quad, faceLimit.
rig_assetGenerate a skeleton and skin weightsA
SPENDS 3D CREDITS. Rigs an existing generated asset — builds a skeleton and skin weights so it can be animated. Takes an asset job that already produced a model. ASYNCHRONOUS: poll get_asset_job, then download_asset. Rig before animating; retargeting an unrigged model wastes the credits. NOTE: the provider task name is taken from published docs and is UNVERIFIED live.
Inputs: assetJobId (required), spec, outFormat.
select_referenceChoose which reference image to reconstructAIdempotent
FREE and local: no network call, no credits. Records which candidate image should be used for 3D reconstruction. Call this after looking at the candidates from generate_asset_reference, and before create_3d_asset.
Inputs: assetJobId (required), candidateId (required).
texture_existing_assetGenerate PBR textures for a mesh you already haveA
SPENDS 3D CREDITS. Applies newly generated PBR materials to an EXISTING mesh, keeping its geometry untouched. Use this when you already own a model and only the materials are missing or wrong — it is cheaper than regenerating and preserves the mesh exactly. Provide the mesh as either modelPath (local .glb/.gltf/.fbx/.obj/.stl) or originalAssetJobId (a previous job from this server whose 3D task produced a model). Direct the material with EITHER prompt OR styleImagePath, not both. ASYNCHRONOUS: poll get_asset_job until "ready", then download_asset. NOTE: providers may only support retexturing meshes that originated from their own prior task; if a provider rejects an uploaded mesh, use originalAssetJobId instead.
Inputs: pbr, spec, prompt, modelPath, +6 more.
validate_game_assetCheck whether an asset is game-readyARead-onlyIdempotent
FREE and fully local: no network call, no credits. Judges a glTF/GLB against a shipping policy and returns a pass/fail verdict with per-check reasons — UVs, normals, tangents where a normal map is bound, triangle budget, material count, texture resolution, power-of-two textures, and bounding-box sanity. Errors fail the asset; warnings are judgement calls and do not. Every threshold is overridable, because an asset failing someone else's house style is not a defect. Pair with normalize_mesh: run this first to see what is wrong, normalize, then run it again to prove the repair.
Inputs: modelPath, assetJobId, requireUVs, maxMaterials, +8 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 20 tools
Each tool has a clearly distinct purpose: generation (reference, variation, full 3D), selection, downloading, inspection, validation, normalization, PBR extraction, sound effects, rigging, animation, retopology, batch operations, and spend reporting. Even similar tools like generate_asset_reference and create_game_prop are differentiated by their roles (raw generation vs. convenience wrapper that stops at candidate selection).
All tool names follow a consistent verb_noun snake_case pattern (e.g., get_asset_job, create_3d_asset, download_asset, validate_game_asset). Minor variation between 'create_' and 'generate_' is acceptable since they reflect different actions (creating a final asset vs. generating intermediate references). The naming is highly predictable.
With 20 tools, the surface is broad but well-scoped for a comprehensive game asset pipeline covering generation, modification, validation, and auxiliary operations. It is on the higher end of typical, but each tool earns its place; no redundant or trivial tools exist.
The toolset covers the full asset lifecycle: generate references, select, create 3D, download, inspect, validate, normalize, extract PBR, rig, animate, retopologize, and batch prepare. Minor gaps include no explicit job deletion or advanced manual editing (e.g., UV sculpting), but these are not core to the stated purpose and can be worked around.
Maintenance
Related MCP Connectors
Generate 3D models, worlds, PBR materials, and coordinated asset packs from AI agents.
Turn prompts into production-ready game assets: animated characters, textures, terrain, VFX, HUD.
Generate 3D models from images or text with AI, then retexture or convert them to STL, 3MF or OBJ.
Generate game-ready 3D models, textures, and audio from natural language, over MCP.
Related MCP Servers
AlicenseAqualityBmaintenanceEnables AI agents to create, manage, and download 3D models, textures, images, rigged characters, and animations through natural conversation.2826,720 npm93MIT
Context3D MCP Serverofficial
AlicenseBqualityDmaintenanceEnables AI-powered 3D model generation from text and images with PBR textures, supporting blockchain authentication and MCP integration.7MIT
Thrixel MCP Serverofficial
AlicenseAqualityBmaintenanceEnables AI agents to create, edit, detail, and download game-ready 3D models through natural conversation, with visual feedback on each result.122,613 PyPIMIT- AlicenseCqualityBmaintenanceEnables AI agents to build game assets in Blender by modeling, lighting, rendering previews to inspect and correct, validating against engine budgets, and exporting to GLB/FBX.126MIT