fast-agent provides a flexible runtime for building LLM-powered agents with first-class support for MCP servers, ACP (Agent Communication Protocol), and reusable Skills. Developers can launch an interactive coding agent session with a single uvx fast-agent-mcp@latest -x command, connect to any MCP server over stdio or HTTP with OAuth, and manage skills via a built-in /skills command. The framework targets the full spectrum of agentic use cases — from exploratory coding sessions to fully automated CI pipelines and evaluation harnesses.
Unlike general-purpose agent frameworks that treat MCP as an afterthought, fast-agent is designed around the protocol from the ground up — supporting Sampling, Elicitations, and other advanced MCP features that most frameworks omit. It supports multiple model providers out of the box (Anthropic Claude, OpenAI Codex, HuggingFace Inference, llama.cpp, and generic local models) and ships with preconfigured packs for common workflows like --pack hf-dev for HuggingFace development or --pack codex for OpenAI Codex-optimized agents. The --smart flag enables automatic subagent routing and compaction strategies for long-running sessions.
fast-agent fits developers who want a coding agent framework that stays close to the MCP standard while remaining lightweight and composable. It works equally well as a daily driver for interactive development (fast-agent --model opus -x --smart) and as an evaluation harness for testing agent behavior across model providers. With 3,700+ GitHub stars, an active Discord community, and regular releases under the Apache 2.0 license, it has become a credible alternative for teams who find LangChain too heavyweight or CrewAI too opinionated for terminal-first coding workflows.
