Show HN: Agent MCP Studio – build multi-agent MCP systems in a browser tab
17 hours ago
- #Agent Development
- #MCP Tools
- #Sandbox Runtime
- Users can create MCP tools by specifying them in plain English, with the LLM generating the corresponding code for a chosen runtime, each with its own sandbox.
- Registration of tools is metadata-only via AST parsing, storing the source code; Python code runs only when a tool is called (JIT), and the LLM sees only enabled tools.
- The system supports Docker deployment for exported servers, DuckDB for querying uploaded CSV or Parquet files, local embeddings for RAG using Xenova/all-MiniLM-L6-v2, and agentic collaboration strategies with personas registered as MCP tools.
- It includes an MCP JSON-RPC 2.0 console for simulating the protocol, a chat backend with options like OpenAI or local WASM LLM, and a bridge to expose browser tools to Claude Desktop via WebSocket.
- Security features include static analysis to block unsafe calls, sandboxed tool execution, network access restricted to an allowlist, and project packs for sharing tools without including credentials.