A Rust desktop hub that gives your agents a memory of their own

What it does

MCPHub Desktop is a Tauri 2 app that manages MCP servers, AI agent Skills, and RAG knowledge bases on your own machine. It’s a Rust rewrite of the upstream web service samanhappy/mcphub — same frontend UI, new backend, plus three personal-use features the upstream never had: Skills management, a RAG knowledge base, and embedded-model management. All processes, configs, keys, and vector data stay local; nothing phones home.

The interesting bit

The RAG pipeline is the star: documents are chunked semantically (plain text by Unicode word/sentence boundaries, Markdown by headings, source code by tree-sitter AST), vectorized with bundled GGUF embedding models running via candle, and stored in lancedb. The knowledge base is then exposed as MCP tools (rag_search, rag_file_create, etc.) — so your agent’s long-term memory is literally a callable tool, not a prompt trick. Skills, meanwhile, install into dozens of agents (Claude Code, Cursor, Cline, Goose…) via symlink or copy, with per-agent install state tracked.

Key highlights

• MCP server processes are hosted by Rust, with Node/UV/Bun runtimes bundled — no preinstalled environment needed
• Hybrid retrieval: vector similarity plus keyword search, with tag filtering
• Chunk sizes auto-adapt per model via each model’s deploy.json, clamped to its context window
• GGUF embedding models (Gemma3, Qwen3, modern-bert, LFM2, nomic-bert-moe) with GPU-first (Metal/CUDA), CPU fallback
• Secrets live in the OS keyring; data in local SQLite ($APPDATA/mcphub.db)

Caveats

• The macOS build is unsigned — expect the “damaged app” dance on first launch
• Windows first-launch crashes occasionally happen after install; a crash.log is written to help debug
• The repo carries a full read-only snapshot of the upstream project (mcphub-origin/), which inflates the clone but keeps the rewrite auditable

Verdict

Worth a look if you juggle multiple MCP servers and agents on one machine and want memory that outlives a chat session. Skip it if you need a team-facing, server-side MCP hub — that’s what the upstream project is for.