Tiny Agents: an MCP-powered agent in 50 lines of code

Source: HuggingFace Date: 2025-04-25 URL: https://huggingface.co/blog/tiny-agents

Summary

Library update and integration tutorial: HF releases @huggingface/mcp-client (NPM), an MCP client on top of their InferenceClient showing a minimal agent loop in ~30 lines of TypeScript. The agent is a while-loop connecting to multiple MCP servers, loading their tools, passing them to an LLM (default: Qwen2.5-72B on Nebius), and looping until completion. Demonstrated with filesystem + Playwright MCP servers. No performance benchmarks.

Implications

Model release cadence (agent reasoning). The reduction of an MCP-capable agent to 30 lines of TypeScript — an LLM call, tool execution, history management, and a termination condition — is an important demystification. Most agent framework complexity is incidental; this shows the essential loop is trivial once MCP handles tool discovery and execution.

HF as open-source ML hub. @huggingface/mcp-client with InferenceClient as the backend means open-weights models accessed through the HF routing layer can now serve as the brain for MCP-compatible agents without any OpenAI dependency. The TypeScript path (vs Python smolagents) expands MCP-based open-weights agents to the web developer ecosystem.