A Python Agent development framework for LLMs featuring native end-to-end MCP support, declarative agent definition, multi-model routing, workflow orchestration, and interactive Coding Agent capabilities.
fast-agent is a Python Agent development framework for LLMs at version 0.6.26, requiring Python ≥ 3.13.5. Its core differentiator is complete end-to-end MCP (Model Context Protocol) support—covering all features including Sampling, Elicitations, and State Transfer across STDIO, SSE, and Streamable HTTP transports. It is the first framework to pass end-to-end tests for all MCP features, and any agent can be exposed as an MCP Server for other clients to consume.
The framework offers declarative agent definition via the @fast.agent() decorator and fastagent.config.yaml configuration. The workflow engine supports four patterns: Chain (sequential), Parallel (fan-out/fan-in), Agents As Tools (orchestrator-worker), and Human Input (human-in-the-loop). MAKER mode reduces error rates in long-chain tasks through multi-sampling with k-vote consensus.
On the model layer, it natively supports 10+ providers including Anthropic, OpenAI, Google (Gemini), Azure, Ollama, Deepseek, Hugging Face, and llama.cpp, with extension to dozens more via the TensorZero gateway. The --model flag supports query override syntax (e.g., claude-sonnet-4-6?web_search=on) for dynamic model behavior adjustment.
For interactive experience, Shell mode enables direct !command execution in the terminal with LSP integration for code awareness. Agent Skills provide progressive capability disclosure via SKILL.md, and pre-built Agent Packs (e.g., hf-dev, codex) enable one-click scenario launch. Enterprise-grade features include built-in OpenTelemetry telemetry (exportable to Hugging Face), optional Privacy Filter (ONNX Runtime-based), and OAuth PKCE authentication.
For protocol extensibility, the fast-agent-acp entry point supports ACP (Agent Client Protocol) depending on agent-client-protocol==0.9.0 and a2a-sdk==0.3.26, with OpenAI Apps SDK (Skybridge) integration. The architecture is built on asyncio strict mode with a CLI layer using Typer + prompt_toolkit + Rich, data validation via Pydantic v2, and a testing hierarchy of unit / integration / e2e / simulated_endpoints.