vrc-udon-mcp-nllabs
# VRC Udon NLLabs
**NeuraLive Labs** · Udon MCP for VRChat
[](https://github.com/NeuraLive-Labs/vrc-udon-mcp-nllabs/releases)
[](LICENSE)
[](https://nodejs.org/)
[](https://modelcontextprotocol.io)
MCP server for VRChat UdonSharp development — exposes the [agent-skills-vrc-udon](https://github.com/niaka3dayo/agent-skills-vrc-udon) knowledge base to AI assistants via the [Model Context Protocol](https://modelcontextprotocol.io).
---
## Documentation
| Language | README | Description |
|----------|--------|-------------|
| **English** | [docs/README.en.md](docs/README.en.md) | Full guide — installation, MCP setup, tools |
| **Español** | [docs/README.es.md](docs/README.es.md) | Guía completa en español |
| **Français** | [docs/README.fr.md](docs/README.fr.md) | Guide complet en français |
| **Japanese** | [docs/README.ja.md](docs/README.ja.md) | Full guide in Japanese |
| **Korean** | [docs/README.ko.md](docs/README.ko.md) | Full guide in Korean |
---
## What is this?
`vrc-udon-mcp-nllabs` (`vrchat-udon-mcp`) is a **stdio MCP server** that indexes, searches, and validates UdonSharp documentation from the `agent-skills-vrc-udon` repository at runtime. No hardcoded docs — the remote repo is the single source of truth.
| | |
|---|---|
| **18 MCP tools** | Search, explain, validate, templates, SDK matrix |
| **Dynamic resources** | Skills, rules, cheatsheets, templates |
| **Live indexing** | MiniSearch with weighted ranking + file watcher |
| **IDE / agent support** | Cursor, Claude Desktop, ChatGPT Desktop, Codex — via MCP + `AGENTS.md` / `CLAUDE.md` |
---
## Quick start
[](https://cursor.com/en/install-mcp?name=vrchat-udon&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyIteSIsImdpdGh1YjpOZXVyYUxpdmUtTGFicy92cmMtdWRvbi1tY3AtbmxsYWJzIl19)
One-click install opens Cursor and adds the portable `npx` MCP entry. Or install from the CLI (merge-safe):
```bash
npx -y github:NeuraLive-Labs/vrc-udon-mcp-nllabs -- install
```
Then **Refresh MCP** in Cursor (CLI install only). Optional: add `--claude` for Claude Desktop.
**Develop from source:**
```bash
git clone https://github.com/NeuraLive-Labs/vrc-udon-mcp-nllabs.git
cd vrchat-udon-mcp
pnpm install
pnpm update-docs && pnpm build-index && pnpm build
pnpm start
```
See [docs/mcp-config.example.json](docs/mcp-config.example.json) and the language READMEs for workspace, `npx`, global install, submodule, and git dependency options.
---
## Agent workflow for accurate Udon code
Works with **Cursor, Claude, Codex, ChatGPT** (any MCP client). Same strong rules everywhere:
| Entry | Role |
|-------|------|
| **[AGENTS.md](AGENTS.md)** | Canonical agent instructions (Codex, Cursor agents, …) |
| **[CLAUDE.md](CLAUDE.md)** | Claude Code / Desktop project context |
| MCP resource `udon://workflow/code` | SSOT summary any MCP client can fetch |
| MCP prompt `udon-code-workflow` | Same workflow as a prompt (Claude-friendly) |
| [`.cursor/rules/udon-mcp.mdc`](.cursor/rules/udon-mcp.mdc) | **Cursor-only** bonus; mirrors `AGENTS.md` |
Flow: **template ? validate ? fix** — consult MCP first (`search_constraints` / `search_networking` / `get_template` or `search_examples`), adapt (never invent networking), then `validate_code` ? `explain_validation` ? fix until clean.
---
## Table of contents
- [Documentation](#documentation)
- [What is this?](#what-is-this)
- [Quick start](#quick-start)
- [Agent workflow for accurate Udon code](#agent-workflow-for-accurate-udon-code)
- [Contributing](#contributing)
- [License](#license)
---
## Contributing
Forks and contributions are welcome. Cualquier fork y ayuda son bienvenidos.
See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.
---
## License
[MIT](LICENSE) — Documentation and skills by [niaka3dayo/agent-skills-vrc-udon](https://github.com/niaka3dayo/agent-skills-vrc-udon). MCP server by [NeuraLive Labs](https://github.com/NeuraLive-Labs) / [MauDevVR](https://github.com/MauDevVR).
TDQS
Scored across 18 tools
The tools are mostly distinct, but the eight search_* variants (documentation, reference, sdk_feature, constraints, networking, examples, best_practice, antipattern) share a common action and could be confused at a glance. Descriptions clearly differentiate them by target, so an agent can usually select correctly, but the overlap prevents a perfect score.
The vast majority of tools follow a consistent verb_noun snake_case pattern (search_*, list_*, read_*, get_*, validate_*, explain_*). The one outlier is sdk_matrix, which is a noun_noun pair without a verb, a minor deviation from the otherwise uniform scheme.
18 tools is on the heavy side, falling in the 16–25 'borderline' range. The domain is specialized and each tool has a real use case, but several search tools could potentially be consolidated into a more generic search with filters, making the set feel slightly over-scoped.
The tool surface covers the full Udon development workflow: discovering skills/rules/templates, reading detailed references, searching for constraints/networking/sdk features, validating code, explaining validation failures, and checking SDK compatibility. No obvious dead ends or missing lifecycle steps remain.