Agent MCP Studio runs multi-agent orchestration in a single browser tab
A browser-only studio for designing and running MCP multi-agent systems ships the entire stack — tool authoring, orchestration, RAG, and code execution — in a single static HTML file powered by WebAssembly, with no backend required.
Score breakdown
Prototype and export production-ready Python MCP servers entirely in-browser — with no infrastructure setup — by leveraging WebAssembly as a free, hard sandbox for safely executing LLM-generated code.
- 01The entire stack runs from a single static HTML file (~5K lines) via WebAssembly — no backend, no server, no Docker required.
- 02Python tools are AST-validated, lazily registered, and JIT-compiled into Pyodide; SQL tools run in DuckDB-WASM in a Web Worker.
- 03Built-in RAG uses Xenova/all-MiniLM-L6-v2 via Transformers.js for on-device embeddings; nothing leaves the browser.
Agent MCP Studio, posted to Hacker News by author stealthtsdb, is a browser-only environment for designing and running MCP-based multi-agent systems. The entire stack — tool authoring, multi-agent orchestration, RAG, and code execution — is delivered as a single static HTML file (~5,000 lines of HTML/CSS/JS) and runs entirely via WebAssembly. Python tools are AST-validated, lazily registered, and JIT-compiled into Pyodide on first call. SQL tools execute in DuckDB-WASM inside a Web Worker. The built-in RAG pipeline uses Xenova/all-MiniLM-L6-v2 via Transformers.js for on-device embeddings. Nothing leaves the browser; closing the tab destroys the entire stack. The WASM boundary acts as the security sandbox, eliminating the need for Docker, per-tenant containers, or a server.
Above the tool layer sits an agentic orchestration system with 10 strategies: Supervisor, Mixture of Experts, Sequential Pipeline, Plan & Execute, Swarm, Debate, Reflection, Hierarchical, Round-Robin, and Map-Reduce.
Above the tool layer sits an agentic orchestration system with 10 strategies: Supervisor, Mixture of Experts, Sequential Pipeline, Plan & Execute, Swarm, Debate, Reflection, Hierarchical, Round-Robin, and Map-Reduce. Users build agent teams visually by dragging tool chips onto persona nodes on a service graph; each persona auto-registers as an MCP tool (`ask_<name>`), and a meta tool `agent_chat(query, strategy?)` is also exposed. A bundled Node bridge (`bridge.js`) speaks stdio to Claude Desktop and WebSocket to the browser tab, turning the browser into an MCP server.
When prototyping is complete, the Export function produces a real Python MCP server: `server.py`, `agentic.py`, `tools/*.py`, `Dockerfile`, `requirements.txt`, and `.env.example`. The exported `agentic.py` is described as a faithful Python port of the browser orchestration logic, so the deployable artifact behaves identically to the prototype. A "Project Packs" feature exports the whole project as a single `.agentpack.json`, auto-detecting required external services (OpenAI, GitHub, Stripe, Anthropic, Slack, Notion, Linear, and others) by scanning tool source for `os.environ.get(...)` calls. The studio also supports a fully offline mode using a local Qwen 1.5 0.5B model running in-browser via Transformers. The author notes open uncertainties including Pyodide cold-start latency on first load and whether all 10 orchestration strategies are necessary in practice.
Key facts
- 01The entire stack runs from a single static HTML file (~5K lines) via WebAssembly — no backend, no server, no Docker required.
- 02Python tools are AST-validated, lazily registered, and JIT-compiled into Pyodide; SQL tools run in DuckDB-WASM in a Web Worker.