threenative-asset-mcp

threenative-asset-mcp

An MCP server for searching, browsing, and downloading free assets from the Fab marketplace, with browser fallback for verification and guarded downloads.

Category
访问服务器

README

threenative-asset-mcp

A Model Context Protocol server for finding 3D assets across Fab, Poly Haven, ambientCG, Smithsonian 3D, Sketchfab, and a curated game-audio catalog spanning Sonniss, Kenney, Tallbeard, Scott Buckley, itch.io, Mixkit, Pixabay, Freesound, OpenGameArt, and Abstraction. It gives AI clients provider-scoped, structured search, asset metadata, category/filter discovery, downloadable file data, and guarded Fab downloads for directly available free files.

The audio tools separate source discovery from verified direct downloads. All ten sources are described with license caveats and official browse pages; only packs with stable official URLs and known license metadata appear in the direct-download catalog. The initial downloadable set is Kenney Interface Sounds, Kenney Music Jingles, and all five Sonniss GDC 2026 archives.

fab_search_assets defaults to free assets. This means Fab reported at least one free or effectively free license; it does not imply every license tier is free. Use fab_get_asset before making license or price claims.

Poly Haven results are CC0 and explicitly labelled Powered by Poly Haven. polyhaven_list_files exposes official download URLs, hashes, sizes, and dependency relationships with pagination and resolution/format filters.

Status: experimental. Fab's /i/* JSON routes are undocumented and can change or restrict automated access. Poly Haven provides a documented public API, but clients must send a unique User-Agent and visibly credit Poly Haven. Sketchfab licenses vary per model and download URLs require a user API token. Review each provider's terms and each asset's license.

Requirements

  • Node.js 20.19 or newer
  • A local environment capable of running Playwright Chromium when Fab requests browser verification
  • No Epic or Fab login is required or automated

Install

An MCP host can launch the published package with:

npx -y threenative-asset-mcp

For a local checkout:

npm ci
npm run browser:install
npm run typecheck
npm test
npm run build
node dist/index.js

Playwright does not download Chromium as part of a normal package install. Run npx -p playwright@1.62.0 playwright install chromium once on the MCP host before relying on the browser fallback.

MCP host configuration

Codex

Add this to ~/.codex/config.toml:

[mcp_servers.assets]
command = "npx"
args = ["-y", "threenative-asset-mcp"]

For a local build:

[mcp_servers.assets]
command = "node"
args = ["/absolute/path/to/threenative-asset-mcp/dist/index.js"]

Claude Desktop

Add a server entry to the Claude Desktop configuration:

{
  "mcpServers": {
    "assets": {
      "command": "npx",
      "args": ["-y", "threenative-asset-mcp"]
    }
  }
}

VS Code

Create .vscode/mcp.json:

{
  "servers": {
    "assets": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "threenative-asset-mcp"]
    }
  }
}

Restart the MCP host after changing its configuration.

MCP tools

Fab:

  • fab_search_assets — searches public listings. priceMode defaults to free; use any or range explicitly for paid results.
  • fab_get_asset — returns normalized public listing details and per-license effective prices.
  • fab_list_filters — returns known public filter labels and slugs, including an explicit warning when the versioned fallback is used.
  • fab_list_limited_time_free — reads only a separately verified curated promotion surface. In production it extracts canonical listing UUIDs from Fab's public /limited-time-free page through the dedicated browser, then resolves them through the normal detail client; it never substitutes general is_free=1 search.
  • fab_download_free_asset — downloads one directly available free file into the dedicated download directory after explicit Fab EULA acknowledgement. It refuses purchase, acquisition, library-only, ambiguous, and unsafe-path flows.

Poly Haven:

  • polyhaven_search_assets — searches HDRIs, textures, and models by text, type, and category, with relevance/popularity/date/name sorting and cursor pagination.
  • polyhaven_get_asset — returns normalized metadata, attributes, authors, dimensions, resolution, and CC0 licensing for one asset.
  • polyhaven_list_categories — returns category labels and counts for one asset type.
  • polyhaven_list_files — returns the official file URLs, sizes, MD5 hashes, and dependency relationships. Use resolution and format to select usable variants; follow nextCursor until absent to retrieve every matching file.

ambientCG:

  • ambientcg_search_assets — searches CC0 materials, HDRIs, substances, decals, atlases, 3D models, images, brushes, terrains, and HDRI elements.
  • ambientcg_get_asset — returns metadata, maps, technique, dimensions, statistics, and CC0 licensing.
  • ambientcg_list_categories — lists typed categories and asset counts.
  • ambientcg_list_files — returns official archives with variant attributes, extensions, URLs, and byte sizes.

Smithsonian 3D:

  • smithsonian_search_assets — searches Open Access models by text, format, quality, owning unit, Draco compression, and glTF orientation compliance.
  • smithsonian_get_asset — groups the file-centric API response into one model summary.
  • smithsonian_list_files — returns direct model files with format, quality, compression, and orientation metadata.

Sketchfab:

  • sketchfab_search_models — anonymously searches public models, defaulting to downloadable results, and preserves author, geometry, archive, and license metadata.
  • sketchfab_get_model — returns public detail and explicit Creative Commons requirements.
  • sketchfab_list_categories — lists public category names and slugs.
  • sketchfab_get_downloads — uses SKETCHFAB_API_TOKEN to retrieve temporary download URLs. It never stores the token or returns it in tool output.

Game audio:

  • audio_list_sources — lists the ten supported audio libraries, best uses, official browse pages, license/attribution cautions, and honest download capability (curated-direct or provider-page).
  • audio_search_assets — searches only the curated packs with stable official direct URLs. Results preserve license, commercial-use, attribution, source, size (when known), and redistribution metadata.
  • audio_download_asset — downloads a catalog asset by ID after acceptLicense: true. It uses an HTTPS host allowlist, validates every redirect, streams with a byte cap, writes atomically without overwrite, and returns the local path, byte size, and SHA-256.

Curated itch.io packs:

  • itch_list_downloads — resolves a fresh no-account download page and lists upload IDs, filenames, sizes, CC0 terms, and pack-specific cautions without exposing the signed page token.
  • itch_download_asset — resolves a fresh 60-second signed file URL and streams the selected upload into guarded storage after acceptLicense: true. Signed URLs are not returned. The initial catalog covers Tallbeard Music Loop Bundle, Quaternius Universal Animation Libraries 1 and 2, Brackeys VFX Bundle, and KayKit Platformer.
  • asset_list_bundle_entries — reads the remote ZIP directory with HTTP byte ranges and returns individual paths and sizes without downloading the archive.
  • asset_download_bundle_entry — range-fetches and extracts one selected file, caches it with a SHA-256 sidecar, and never downloads unrelated bundle files.
  • asset_list_bundle_animations — range-fetches only an aggregate GLB and lists its named animation clips. For Quaternius, it automatically prefers the standard non-root-motion GLB.
  • asset_download_bundle_animation — exports one named animation as a valid animation-only GLB, removing unrelated clips, meshes, materials, and textures. The cached aggregate GLB is reused across requests.

Unified source and download routing:

  • asset_list_sources — returns only agent-ready sources by default across 3D, textures, HDRIs, animations, VFX, 2D, UI, icons, fonts, and audio. An agent-ready source has an MCP download tool and requires no manual browser, login, checkout, donation prompt, or paywall. Pass agentReadyOnly: false to inspect the broader research directory, including package-manager, Git, and provider-page sources that are not yet integrated.
  • asset_search_sources — filters that directory by text, category, access mode, and license tag (cc0, cc-by, mit, and others), while preserving the same agent-ready-only default.
  • asset_download_file — streams a direct URL previously returned by polyhaven_list_files, ambientcg_list_files, smithsonian_list_files, or the Game-icons.net bulk archive or Kenney Particle Pack entry into guarded local storage. Provider hosts and URL shapes are allowlisted, redirects are revalidated, existing files are never overwritten, and the result includes SHA-256.

Recommended agent flow:

  1. Call asset_search_sources with the requested category/query. Its default result set is guaranteed to contain only agent-ready sources.
  2. Call the returned searchTool or detailTool when present.
  3. Resolve variants with the returned filesTool.
  4. Call the returned downloadTool with acceptLicense: true.
  5. For aggregate Quaternius animation libraries, skip whole-pack download: itch_list_downloadsasset_list_bundle_animationsasset_download_bundle_animation.

This is intentionally a short MCP tool chain rather than a fake universal URL: each provider keeps its real search/variant semantics, while source routing and the no-manual-flow guarantee stay uniform.

All discovery tools are read-only. Download tools write only within their dedicated local directories and never purchase, add to cart or library, wishlist, sign in, or overwrite an existing download. Audio packs remain subject to their source license; raw redistribution is not implied by download.

Configuration

Variable Default Purpose
FAB_DIRECT_TIMEOUT_MS 20000 Direct JSON request timeout.
FAB_BROWSER_TIMEOUT_MS 30000 Dedicated browser request timeout.
FAB_BROWSER_MANUAL_TIMEOUT_MS 10000 Headed-mode grace period for visible verification.
FAB_BROWSER_HEADLESS true Set to 0 temporarily for manual verification.
FAB_BROWSER_PROFILE_DIR OS state directory under threenative-asset-mcp/fab-browser-profile MCP-owned Fab browser state.
FAB_DOWNLOAD_DIR ~/Downloads/threenative-asset-mcp/fab Dedicated directory for Fab free-file downloads.
FAB_MAX_DOWNLOAD_BYTES 2147483648 Maximum accepted download size in bytes.
FAB_DOWNLOAD_TIMEOUT_MS 600000 Total timeout for one file download.
FAB_CURL_IMPERSONATE auto-detected on PATH curl-impersonate wrapper override; 0/off disables.
FAB_MIN_REQUEST_INTERVAL_MS 1000 Minimum spacing between direct upstream requests.
FAB_LOG_LEVEL warn debug, info, warn, or error.
FAB_LOG_QUERIES false Set to 1 only if query text may be written to logs.
SKETCHFAB_API_TOKEN unset User token for temporary Sketchfab download URLs.
AUDIO_DOWNLOAD_DIR ~/Downloads/threenative-asset-mcp/audio Dedicated directory for curated audio downloads.
AUDIO_MAX_DOWNLOAD_BYTES 10737418240 Maximum accepted bytes per audio archive (10 GiB).
AUDIO_DOWNLOAD_TIMEOUT_MS 1800000 Total timeout for one audio download (30 minutes).
ASSET_DOWNLOAD_DIR ~/Downloads/threenative-asset-mcp/assets Dedicated directory for direct provider downloads.
ASSET_MAX_DOWNLOAD_BYTES 10737418240 Maximum accepted bytes per provider file (10 GiB).
ASSET_DOWNLOAD_TIMEOUT_MS 1800000 Total timeout for one provider download (30 minutes).

Direct requests are spaced at least FAB_MIN_REQUEST_INTERVAL_MS apart. Only HTTP 429, 502, 503, and 504 are retried, at most twice, with backoff and Retry-After support. Challenges, access denial, invalid input, missing listings, and schema drift are never retried by the transport; a curl-impersonate challenge response is additionally retried inside the impersonation wrapper with longer, jittered waits before the browser fallback is engaged.

Browser-fingerprint TLS (curl-impersonate)

Fab's /i/* JSON routes sit behind Cloudflare bot management that challenges Node's default TLS fingerprint — including plain fetch from this MCP and headless Chromium — while real browser fingerprints pass. When a curl-impersonate wrapper (for example curl_chrome146) is available on the host PATH, the server performs all anonymous JSON reads through it and no browser is needed for search, detail, or download resolution. Set FAB_CURL_IMPERSONATE=0 to force the old behavior (plain Node fetch plus the Playwright fallback), or point it at a specific wrapper binary.

Downloads of free files resolve entirely through the anonymous JSON contract when possible: listing detail → asset-formats/{format} file listing → download-info signed URL → guarded file write. The signed distribution URL is validated against an exact Epic distribution-host allowlist before use. Only when the direct path is challenged does the server fall back to the guarded browser click flow below.

Process-local cache TTLs are five minutes for search, fifteen minutes for listing details, six hours for taxonomy data, and ten minutes for promotions. The shared LRU is capped at 500 entries and is cleared on process exit.

Poly Haven API behavior

Poly Haven requests go only to https://api.polyhaven.com, with the required threenative-asset-mcp User-Agent. Asset lists are cached for 15 minutes and details, categories, and file trees are cached for up to one hour. Returned file URLs are accepted only from https://dl.polyhaven.org.

The live API is free for personal and commercial use, but use of the API requires a visible Poly Haven credit. The assets themselves are CC0. This MCP includes provider, license, and attribution fields so downstream clients can preserve that distinction. See the official API page and API documentation.

Other provider behavior

ambientCG uses its anonymous, read-only v3 API. Its files are CC0 and its API exposes searchable metadata, categories, and downloadable variants.

Smithsonian uses the anonymous Smithsonian 3D file-search API. Files exposed by that API are part of Smithsonian Open Access; the MCP retains direct source URLs and format/quality metadata.

Sketchfab public search, categories, and model detail do not require a token. The download endpoint requires a token belonging to the user, configured through SKETCHFAB_API_TOKEN. The token is sent only in the Sketchfab Authorization header, is never logged or persisted, and is not included in MCP responses. Sketchfab models use different Creative Commons licenses; always inspect license.requirements before use.

Audio direct downloads are catalog-ID based; the MCP does not accept arbitrary URLs. Official Kenney downloads are restricted to kenney.nl and Sonniss GDC downloads to downloads.sonniss.com, including redirect revalidation. Sources without a stable, verified direct contract remain discoverable as provider-page instead of being falsely presented as one-click downloads.

The generic direct downloader is intentionally narrower than an arbitrary URL fetcher. It accepts only official Poly Haven, ambientCG, Smithsonian, Game-icons.net, and curated Kenney URL contracts. The itch.io downloader uses fresh signed mirror URLs internally but never exposes them. Sketchfab signed downloads are exposed through sketchfab_get_downloads but are not persisted by the generic downloader because their temporary CDN hosts vary and require the user's token-backed session. Provider-page-only sources stay provider-page-only until a stable, license-safe download contract is verified.

Dedicated browser profile and privacy

When a direct anonymous request receives a Cloudflare challenge, the server may open Playwright Chromium with a dedicated MCP-owned profile. It never attaches to, copies, or reads the user's normal Chrome/Edge/Chromium profile, cookie database, local storage, passwords, or Epic session.

If the tool returns FAB_BROWSER_ATTENTION_REQUIRED, run the same MCP command once with FAB_BROWSER_HEADLESS=0. The MCP opens its dedicated Fab homepage at startup, so complete any visible verification before calling the tool. A tool call allows an additional FAB_BROWSER_MANUAL_TIMEOUT_MS grace period, then returns FAB_BROWSER_ATTENTION_REQUIRED rather than exceeding typical MCP client timeouts. Close the MCP process after verification and return to headless mode. The server does not solve or bypass challenges.

Browser process startup is capped at ten seconds. On a host without a working graphical session, headed mode returns FAB_UPSTREAM_UNAVAILABLE instead of hanging an MCP call; run the manual release gate on a graphical host.

To clear browser-owned Fab state, stop every threenative-asset-mcp process and move only the dedicated directory reported by your configuration out of service. The default on Linux can be cleared recoverably with:

mv -- "${XDG_STATE_HOME:-$HOME/.local/state}/threenative-asset-mcp/fab-browser-profile" \
  "${XDG_STATE_HOME:-$HOME/.local/state}/threenative-asset-mcp/fab-browser-profile.cleared"

Do not point FAB_BROWSER_PROFILE_DIR at a normal browser profile. The server rejects known normal-profile locations.

The MCP has no analytics or remote telemetry. Application logs are structured JSON written only to stderr; stdout is reserved for MCP JSON-RPC. Search query text is omitted from logs unless FAB_LOG_QUERIES=1. Raw upstream bodies, headers, cookies, tokens, stack traces, and browser profile paths are not logged or returned to the model.

Troubleshooting

FAB_CHALLENGE or FAB_BROWSER_ATTENTION_REQUIRED : Fab asked for browser verification. First check whether a curl-impersonate wrapper is installed (FAB_LOG_LEVEL=info logs fab_impersonate_enabled when active). Otherwise use the headed dedicated-profile step above. If verification continues to fail, stop; do not copy a signed-in browser session.

FAB_RATE_LIMITED : Wait for retryAfterSeconds when present. The MCP already applied its bounded retries.

FAB_UPSTREAM_CHANGED : Fab's undocumented response changed. Re-run the sanitized contract probe and update normalization and fixtures before continuing.

FAB_UPSTREAM_UNAVAILABLE : Fab is unavailable, or a currently unverified discovery contract was intentionally disabled.

The MCP host shows no tools : Build first, confirm the configured path is absolute, and run npm run inspect. Logs belong on stderr; any non-JSON stdout is a bug.

Verification

npm ci
npm run browser:install
npm run typecheck
npm test
npm run build
npm pack --dry-run
npm run inspect
npm run test:providers:live

Live checks are opt-in because they contact Fab:

npm run test:live

This runs the real anonymous search, cursor, detail, and dedicated-browser contract probe. It exits nonzero when Fab challenges the clean browser or the required contract cannot be verified. Live verification must remain anonymous, concurrency-one, capped and paced. It must never acquire, purchase, wishlist, download, or automatically solve a challenge.

npm run test:providers:live launches the compiled stdio MCP and exercises live search, detail, categories, and file discovery for ambientCG, Smithsonian 3D, and Sketchfab. If SKETCHFAB_API_TOKEN is configured it also verifies the authenticated download endpoint; otherwise it verifies the explicit authentication-required response.

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选