unity-api-mcp
MCP server providing accurate Unity API documentation to prevent AI agents from hallucinating signatures, namespaces, and deprecated usage.
README
unity-api-mcp
<!-- mcp-name: io.github.Codeturion/unity-api-mcp -->
MCP server that gives AI agents accurate Unity API documentation. Prevents hallucinated signatures, wrong namespaces, and deprecated API usage.
Supports Unity 6 (one database per minor stream), Unity 2023, and Unity 2022 LTS. Works with Claude Code, Cursor, Windsurf, or any MCP-compatible AI tool. No Unity installation required. See the supported versions. New Unity releases are detected and built automatically every week.
Quick Start
Add to your MCP config (.mcp.json, mcp.json, or your tool's MCP settings), setting UNITY_VERSION to match your project:
{
"mcpServers": {
"unity-api": {
"command": "uvx",
"args": ["unity-api-mcp"],
"env": {
"UNITY_VERSION": "6000.3"
}
}
}
}
Valid values: a Unity 6 stream like "6000.3", or "6", "2023", "2022".
On first run the server downloads the correct database (~20-30 MB) to ~/.unity-api-mcp/.
How It Works
- Version detection. The server figures out which Unity version to serve:
| Priority | Source | Example |
|---|---|---|
| 1 | UNITY_VERSION env var |
"2022", "6", "6000.3", or "6000.3.8f1" |
| 2 | UNITY_PROJECT_PATH |
Reads ProjectSettings/ProjectVersion.txt, maps 2022.3.62f1 to "2022", 6000.3.8f1 to "6000.3" |
| 3 | Default | "6" |
-
Database download. If the database for that version isn't cached locally, it downloads from GitHub. Unity 6 minor streams (
6000.0,6000.3,6000.5, …) get their own per-stream database, falling back to the generic6database when a stream database isn't published. Cached databases are freshness-checked against the release on startup, so weekly rebuilds reach existing installs automatically. -
Serve. All tool calls query the version-specific SQLite database. Every query returns in <15ms.
Each version has its own database with the correct signatures, deprecation warnings, and member lists for that release.
Tools
| Tool | Purpose | Example |
|---|---|---|
search_unity_api |
Find APIs by keyword | "Tilemap SetTile", "async load scene" |
get_method_signature |
Exact signatures with all overloads | UnityEngine.Physics.Raycast |
get_namespace |
Resolve using directives |
"SceneManager" -> using UnityEngine.SceneManagement; |
get_class_reference |
Full class reference card | "InputAction" -> all methods/fields/properties |
get_deprecation_warnings |
Check if an API is obsolete | "WWW" -> Use UnityWebRequest instead |
Coverage
All UnityEngine and UnityEditor modules, plus packages parsed from C# source: Input System, Addressables, uGUI (incl. TextMeshPro on Unity 6), AI Navigation, and Netcode. ~42,500 records per Unity 6 database, ~500 deprecation warnings each.
Full version list: db-v1 release page. CI regenerates that table on every build. New Unity patches are detected and built automatically every Monday.
Does not cover third-party assets (DOTween, VContainer, Newtonsoft.Json). For those, rely on project source.
Benchmarks
Measured, not promised: 25 research questions across 3 testbeds, answered by 3 agent configs, every answer judged against ground truth verified in the source beforehand. The full harness lives in docs/benchmark/ and re-runs with one command.
<picture> <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/Codeturion/unity-api-mcp/master/docs/images/benchmark-accuracy-dark.png"> <img alt="Answer quality by agent config, judged against verified ground truth" src="https://raw.githubusercontent.com/Codeturion/unity-api-mcp/master/docs/images/benchmark-accuracy-light.png"> </picture>
| Config | Correct | Partial | Wrong | Hallucinated APIs |
|---|---|---|---|---|
| MCP + targeted Read | 24/25 | 1 | 0 | 0 |
| Skilled (Grep+Read) | 20/25 | 3 | 2 | 1 |
| Naive (full Reads) | 19/25 | 3 | 3 | 1 |
The one reproduced hallucination is instructive. Asked to list the overloads of SceneManager.LoadSceneAsync, both non-MCP agents invented single-parameter LoadSceneAsync(string) and LoadSceneAsync(int) overloads that do not exist. Code written against them does not compile. The MCP agent returned the exact four real overloads.
Why correctness and not token savings? Agentic tools are good at code search now. Claude Code has shipped a Grep tool from the start, and current models search first and then read a narrow line range, so raw token use was comparable across all configs in our runs. But exact overloads, namespaces, and deprecations are not in your project files at all. An agent without MCP can only infer them from usage examples, and when it infers wrong you pay with a broken build.
<details> <summary>Methodology</summary>
- 3 testbeds: a real Unity 6 game project (11 questions), pure Unity API lookups (8), and Unity Input System package source with 2,700 to 4,600 line files (6)
- 3 configs, same model and turn limit: MCP tools + Grep/Read, Grep/Read only, Read only
- Ground truth verified by reading the source before any runs; answers judged by a separate model session against that ground truth; token usage taken from API usage fields
- Run it yourself:
python docs/benchmark/run.py --project <unity-project-path>(results from July 2026; agent behavior moves, so re-run before quoting)
</details>
CLAUDE.md Snippet
Add this to your project's CLAUDE.md (or equivalent instructions file). This step is important. Without it, the AI has the tools but won't know when to reach for them.
## Unity API Lookup (unity-api MCP)
Use the `unity-api` MCP tools to verify Unity API usage instead of guessing. **Do not hallucinate signatures.**
| When | Tool | Example |
|------|------|---------|
| Unsure about a method's parameters or return type | `get_method_signature` | `get_method_signature("UnityEngine.Tilemaps.Tilemap.SetTile")` |
| Need the `using` directive for a type | `get_namespace` | `get_namespace("SceneManager")` |
| Want to see all members on a class | `get_class_reference` | `get_class_reference("InputAction")` |
| Searching for an API by keyword | `search_unity_api` | `search_unity_api("async load scene")` |
| Checking if an API is deprecated | `get_deprecation_warnings` | `get_deprecation_warnings("FindObjectOfType")` |
**Rules:**
- Before writing a Unity API call you haven't used in this conversation, verify the signature with `get_method_signature`
- Before adding a `using` directive, verify with `get_namespace` if unsure
- Covers: all UnityEngine/UnityEditor modules, Input System, Addressables, uGUI/TextMeshPro, AI Navigation, Netcode
- Does NOT cover: DOTween, VContainer, Newtonsoft.Json (third-party)
Setup Details
<details> <summary>Auto-detect version from project path</summary>
Instead of setting UNITY_VERSION, you can point to your Unity project. The server reads ProjectSettings/ProjectVersion.txt automatically:
{
"mcpServers": {
"unity-api": {
"command": "uvx",
"args": ["unity-api-mcp"],
"env": {
"UNITY_PROJECT_PATH": "/path/to/your/unity-project"
}
}
}
}
</details>
<details> <summary>Alternative installation methods</summary>
Using pip install:
pip install unity-api-mcp
{
"mcpServers": {
"unity-api": {
"command": "unity-api-mcp",
"args": [],
"env": {
"UNITY_VERSION": "6000.3"
}
}
}
}
</details>
<details> <summary>Environment variables</summary>
| Variable | Purpose | Example |
|---|---|---|
UNITY_VERSION |
Unity version to serve | 6000.3, 6000.3.8f1, 6, 2023, 2022 |
UNITY_PROJECT_PATH |
Auto-detect version from project | F:/Unity Projects/my-project |
UNITY_INSTALL_PATH |
Override Unity install path (for ingest only) |
D:/Unity/6000.3.8f1 |
</details>
<details> <summary>Building databases locally</summary>
If you want to build a database from your own Unity installation instead of downloading:
# Install with ingest dependencies
pip install unity-api-mcp[ingest]
# Windows
python -m unity_api_mcp.ingest --unity-version 6000.3 --unity-install "D:/Unity/6000.3.8f1" --project "F:/Unity Projects/MyProject"
# macOS
python -m unity_api_mcp.ingest --unity-version 6000.3 --unity-install "/Applications/Unity/Hub/Editor/6000.3.20f1" --project "/path/to/UnityProject"
# Legacy versions
python -m unity_api_mcp.ingest --unity-version 2022 --unity-install "D:/Unity/2022.3.62f1"
Databases are written to ~/.unity-api-mcp/unity_docs_{version}.db by default.
</details>
<details> <summary>AI-Assisted Setup</summary>
If an AI agent is setting this up for you:
Add
unity-api-mcpto my MCP config usinguvxwithUNITY_VERSIONset to match my project, append the CLAUDE.md snippet from the README, and verify withget_namespace("SceneManager").
</details>
<details> <summary>Project structure</summary>
unity-api-mcp/
├── src/unity_api_mcp/
│ ├── server.py # MCP server (5 tools)
│ ├── db.py # SQLite + FTS5 database layer
│ ├── version.py # Version detection + DB download
│ ├── xml_parser.py # Parse Unity XML IntelliSense files
│ ├── cs_doc_parser.py # Parse C# doc comments from package source
│ ├── unity_paths.py # Locate Unity install + package dirs
│ └── ingest.py # CLI ingestion pipeline
└── pyproject.toml
Databases are stored in ~/.unity-api-mcp/ (downloaded on first run).
</details>
Troubleshooting
| Problem | Fix |
|---|---|
| "Could not download Unity X database" | Check internet connection. Or build locally: python -m unity_api_mcp.ingest --unity-version 2022 |
| Wrong API version being served | Set UNITY_VERSION explicitly. Check stderr: unity-api-mcp: serving Unity <version> API docs |
| Server won't start | Check python --version (needs 3.10+). Check path: which unity-api-mcp or where unity-api-mcp |
| Third-party packages return no results | DOTween, VContainer, Newtonsoft.Json are not indexed (third-party, not Unity packages) |
See Also
unreal-api-mcp: same concept for Unreal Engine (C++), with weekly auto-built databases per UE version.
Contact
Need a custom MCP server for your engine or framework? I build MCP tools that cut token waste and prevent hallucinations for AI-assisted game development. If you want something similar for your team's stack, reach out.
fuatcankoseoglu@gmail.com
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。