Overview
fast-agent is a flexible way to work with LLMs from the command line: as a coding agent, a development toolkit, or a platform for building and evaluating agent workflows. It has deep MCP support, works with ACP clients, and lets you define agents declaratively so you can focus on prompts and the MCP servers they use.
Key capabilities
- Interactive shell: chat, run shell commands with
!, and connect MCP servers with/connect. - Skills: manage agent skills (SKILL.md) from a registry.
- MCP depth: sampling, elicitations, MCP Apps and transport diagnostics.
- ACP: run any fast-agent setup inside ACP clients.
- Wide model support: Anthropic, OpenAI, Google, local llama.cpp and more.
Tradeoffs
- CLI-first; there's no full graphical app.
- A broad feature set means a learning curve.
- Smaller community than the big agent frameworks.
Ideal for
- Developers building and testing MCP servers
- Evaluating agents against several models
- Terminal-first agent workflows
Not ideal for
- Non-technical users
- Teams wanting a visual builder
Getting started
- Install uv, then try it without installing:
uvx fast-agent-mcp@latest -xopens an interactive session with shell support. - Install it properly with
uv tool install -U fast-agent-mcp, which also sets up shell aliases. - Connect an MCP server from inside the session with
/connect, and manage skills with/skills. - For local models, run
fast-agent model llamacppto generate the right configuration.
Technical notes
- Install:
uv tool install -U fast-agent-mcp - Docs: fast-agent.ai
- License: Apache-2.0
- Source: evalstate/fast-agent
- GitHub stars: 3,964 (as of 2026-11-12)