Markgatcha/universal-mcp-toolkit

English · 简体中文 · Español · 日本語 · 한국어 · Français · Deutsch · Português · Русский · हिन्दी · العربية
The canonical open-source monorepo for production-ready Model Context Protocol servers.
If you have ever wanted one place to find great MCP servers for GitHub, Slack, Notion, databases, cloud platforms, research sources, and local files without stitching together a dozen half-finished repos, this is it.
🔗 Part of the AI Trio
universal-mcp-toolkit is one of three sibling projects that compose into a complete agent memory + tooling stack:
| Project | Role |
| --- | --- |
| **universal-mcp-toolkit** | MCP protocol, server registry, and tool routing |
| **memos** | Graph-based persistent memory across agent sessions |
| **llm-guardian** | Token-cost guardian that compresses prompts and injects MemOS memory slices |
Together they cover transport + tools (UMT), memory + persistence (MemOS), and LLM inference cost control (llm-guardian). The MemOS MCP adapter is published as @mem-os/sdk and pairs directly with UMT's link memos command.
⚡ Quick Start
The fastest way to get going:
For a guided setup walkthrough, open docs/getting-started.html. For deterministic composition, see docs/workflows.md.
Or install globally:
Or install from CodeGuilds — the community registry for AI developer tools:
Why this exists
The MCP ecosystem is exploding, but the developer experience is still fragmented.
- Most repos solve one narrow integration.
- Many servers stop at a demo-quality tool or two.
- Transport support, auth handling, docs, and packaging are wildly inconsistent.
- There is no obvious reference implementation that shows how a serious MCP monorepo should feel.
universal-mcp-toolkit fixes that with one opinionated, high-quality Turborepo:
- 31 production-focused MCP servers
- One shared strict-mode TypeScript core
- One polished CLI for install, config, run, and diagnostics
- Consistent Zod validation, structured errors, and pino logging
- Stdio plus HTTP+SSE support across the toolkit
- Discovery-friendly .well-known/mcp-server.json server cards
What makes this worth starring
- **Tool discovery** — umt tools list finds any MCP tool across 27+ servers by name or category
- **Deterministic workflows** — umt workflow validate|run executes reviewable, versioned JSON workflows with strict input and step references
- **Server composition** — umt compose remains available for quick two-step output piping
- **Standards-aligned discovery** — umt discover --registry reads official MCP Registry-compatible endpoints alongside local and well-known discovery
- **Enforced tool boundaries** — bridge allowlists and RBAC policies are checked before cache lookup, reconnect, or remote execution
- **TTL + LRU caching** — MCPFunctionCallingBridge caches results by tool+args, avoiding redundant API calls
- **Health monitoring** — transport failures feed the circuit breaker without treating ordinary tool validation errors as connection failures
- **Buffered result chunks** — callToolStreaming() exposes completed results through an async chunk iterator; it does not claim protocol-level streaming
- **Multi-server sessions** — Session class orchestrates tools across multiple MCP servers with parallel calls
- **Lazy plugin loading** — server packages loaded on-demand, not statically bundled
- **Lazy tool registration** — registerLazyTool defers expensive initialization until a tool is actually called
- **Structured error context** — bridge errors include tool name, args, and error type for easier debugging
- **Type-safe chaining** — callToolChain with ToolChain types enables compile-time-checked tool pipelines
- Real developer utility right now
- Great default ergonomics for Claude Desktop, Cursor, and local workflows
- A single architecture you can learn once and extend everywhere
- Strong package hygiene with exports maps, keywords, build scripts, and test hooks
- A repo designed to be both a product and a reference implementation
The short version
| Category | What you get |
| --- | --- |