fast-agent is a Python framework and command-line tool for building and interacting with LLM agents and multi-agent workflows over the Model Context Protocol. It serves as an interactive coding agent, a development toolkit, and an evaluation or workflow platform: a single command opens an interactive session with shell access, and the same agent definitions run as scripted automations or as MCP servers.
Agents are declared with decorators. @fast.agent takes an instruction, a model, MCP servers, and options such as history and human input. Workflow decorators compose those agents into chains, parallel fan-out and fan-in, evaluator-optimizer loops, LLM routers, orchestrators that plan and delegate, agents-as-tools hierarchies, and MAKER, which samples a worker repeatedly until one answer leads by a k-vote margin. Every agent or workflow responds to .send() or opens an interactive prompt, and you can chat with individual components before, during, and after a workflow run.
MCP support covers the full feature set, including sampling, elicitations, roots, MCP Apps, and the OpenAI Apps SDK, with end-to-end tests. HTTP and SSE servers use OAuth by default with PKCE and keyring-backed token storage, and a transport diagnostics view shows Streamable HTTP usage. Model providers include Anthropic, OpenAI, and Google natively, plus Azure, Ollama, DeepSeek, Hugging Face inference providers, llama.cpp, and many others through TensorZero, and model strings accept query overrides such as ?reasoning=low or ?web_search=on.
Features
- Interactive CLI: prompt_toolkit-based terminal with completions, in-terminal menus, streaming output, shell commands via !, and /skills and /connect commands
- Declarative agents: @fast.agent with instruction, model, MCP servers, history, request parameters, and optional human input
- Workflow patterns: chain, parallel, evaluator-optimizer, router, orchestrator (full or iterative planning), agents as tools, and MAKER k-voting
- Full MCP feature coverage: sampling, elicitations, roots, prompts, MCP Apps, and Streamable HTTP transport diagnostics
- MCP OAuth: PKCE flow with a local callback server, keyring token persistence, and per-server opt-out
- Server mode: expose any agent as an MCP server over HTTP
- ACP clients: run inside any Agent Client Protocol client through fast-agent-acp
- Card packs: packaged starter agents such as hf-dev, codex, and analyst installed with --pack
- Agent Skills: manages SKILL.md skills from a registry covering LSP, hooks, compaction strategies, and automation
- Multimodal and structured output: PDF and vision inputs, structured outputs, and Passthrough and Playback LLMs for testing glue code
- Provider coverage: Anthropic, OpenAI, Google, Azure, Ollama, DeepSeek, Hugging Face, llama.cpp, and others via TensorZero
