unity-asset-reference-mcp
Indexes Unity project assets into a SQLite reference graph and provides an MCP server for querying asset dependencies, references, unused assets, and broken references.
README
unity-asset-reference-mcp
Index a Unity project's assets into a reusable SQLite reference graph, then query it three ways: a CLI, an MCP server (for Claude and other agents), and a local web viewer. Unity-only.
- Impact analysis — "what references this material?" before you change it
- Unused-asset detection — orphans nothing loads (Addressables-aware)
- Dependency tracing — the full closure a scene/prefab pulls in
- Broken-reference detection — refs to deleted/missing assets
- Reusable artifact — a plain
.sqliteany tool can open; no lock-in
It reads Unity's own serialization (.meta GUIDs + YAML {fileID, guid} refs),
so the graph is exact — no heuristics, no AST. Requires Asset Serialization:
Force Text (Unity's default for version control).
Requirements
- Node.js ≥ 20
- A Unity project using Force Text serialization
better-sqlite3builds via prebuilt binaries during install (native module)
Install
npm install -g unity-asset-reference-mcp
This puts three commands on your PATH:
| Command | What |
|---|---|
unity-asset-reference-mcp-index |
the CLI indexer |
unity-asset-reference-mcp |
the MCP server |
unity-asset-reference-mcp-web |
the web viewer |
Without a global install (npx): the MCP server's bin name matches the package,
so npx -y unity-asset-reference-mcp … works directly. For the other two bins,
name the package with -p:
npx -y unity-asset-reference-mcp --project /path/to/UnityProject # MCP server
npx -y -p unity-asset-reference-mcp unity-asset-reference-mcp-index index /path # indexer
npx -y -p unity-asset-reference-mcp unity-asset-reference-mcp-web --db <index.db> # viewer
1. Index a project
Builds <project>/.asset-memory/index.db.
unity-asset-reference-mcp-index index /path/to/UnityProject --force
Add this to your Unity project's .gitignore — ignore the live index, but
commit the config and (optional) shared snapshot:
# unity-asset-reference-mcp: ignore the live index, keep config + shared snapshot
.asset-memory/index.db
.asset-memory/index.db-*
.asset-memory/*.building-*
2. MCP server (works with any MCP client)
This is a standard stdio MCP server — it works with any MCP-compatible host: Claude Code/Desktop, Cursor, Windsurf, Cline, VS Code (Copilot agent), Zed, and others. Nothing is Claude-specific; only where you put the config differs.
Generic config (Claude Desktop, Cursor, Windsurf, Cline, and most hosts use
this mcpServers shape):
{
"mcpServers": {
"unity-asset-graph": {
"command": "npx",
"args": ["-y", "unity-asset-reference-mcp", "--project", "/path/to/UnityProject"]
}
}
}
Where that config lives, per host:
| Host | Config location |
|---|---|
| Claude Code | claude mcp add unity-asset-graph -- npx -y unity-asset-reference-mcp --project /path/to/UnityProject (or .mcp.json in the project) |
| Claude Desktop | claude_desktop_config.json |
| Cursor | .cursor/mcp.json (project) or ~/.cursor/mcp.json (global) |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
| Cline | cline_mcp_settings.json |
| VS Code (Copilot) | .vscode/mcp.json — uses the key servers instead of mcpServers |
Prerequisites: Node ≥ 20 on PATH (for npx), and a Unity project on Force
Text serialization. You do not need to pre-index — call the index_project
tool once from the agent and it builds <project>/.asset-memory/index.db; the
read tools return a clear no-index error until you do.
Tools exposed: index_project, index_status, get_dependencies,
find_references, find_unused_assets, trace_path, search_assets,
get_overview.
3. Web viewer
# server flavor — serves the viewer + a JSON API over the index
unity-asset-reference-mcp-web --db /path/to/UnityProject/.asset-memory/index.db
# open http://localhost:7777
There is also a static flavor: open dist/web/public/viewer.html in a
browser and pick a .db — it runs the same queries entirely in-browser (WASM
SQLite), no server. Both share one query layer.
Team sharing (snapshots)
Instead of every teammate re-indexing from scratch, commit a compressed
snapshot and let them restore it — the same idea as codebase-memory's
.codebase-memory/.
# after indexing, export a shareable snapshot (or pass --snapshot to `index`)
unity-asset-reference-mcp-index snapshot /path/to/UnityProject
# -> .asset-memory/index.db.br (brotli, ~85% smaller) + artifact.json + .gitattributes
# a teammate who clones + pulls restores the live index with no re-index:
unity-asset-reference-mcp-index restore /path/to/UnityProject
The MCP server and web viewer auto-restore from a snapshot on first use if the live index is missing, so a fresh clone "just works". Commit these:
.asset-memory/config.json # per-project settings (below)
.asset-memory/index.db.br # compressed shared index
.asset-memory/artifact.json # snapshot metadata (schema, commit, counts)
.asset-memory/.gitattributes # marks the blob binary + merge=ours
artifact.json records the git commit and counts the snapshot was built at, so
you can tell when it's stale and re-run index --snapshot.
Configuration
Optional per-project .asset-memory/config.json (see
docs/product/configuration.md):
{
"unused": { "addressableRoots": "auto" },
"scan": { "ignore": ["**/ThirdParty/**", "*.bak"], "ignoreDefaults": true }
}
unused.addressableRoots(auto|on|off) — count Addressable entries as roots for unused detection (query-time;auto= on if the project uses Addressables). Overridable per call.scan.ignore/scan.ignoreDefaults— extra ignore globs (index-time).
The SQLite artifact
Open .asset-memory/index.db with any SQLite tool. Core tables: assets
(nodes), edges (references), unresolved_refs (broken refs),
addressable_entries, index_meta. Schema:
docs/product/asset-graph-model.md.
Known limitations
- Code-based refs not tracked:
Resources.Load("path")and hard-coded Addressable address strings in C# aren't scanned yet, sofind_unusedoutput is candidates — verify against your loading code. - Asset-level granularity: edges are asset→asset (not per-GameObject/fileID).
- Incremental re-index is by mtime; use
--forcefor a fully consistent rebuild.
Development
npm install
npm test # vitest
npm run build # tsc + copy web assets to dist/
This repo also uses an internal agent Harness for development process (see docs/HARNESS.md); it is not part of the shipped package.
License
MIT — see LICENSE.
推荐服务器
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。
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。
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。
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。
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。
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。
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。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
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。