Fast Agent 🤖 Agent Open source

What it does

fast-agent is a Python framework — not a packaged end-user app — for building, running and evaluating AI agents. Its own README describes it as "a flexible way to interact with LLMs, excellent for use as a Coding Agent, Development Toolkit, Evaluation or Workflow platform." Agents are declared with simple Python decorators and can be chained, parallelized, or composed into routing/orchestration workflows.

What is inside

• Declarative @fast.agent syntax for defining agents and instructions
• Built-in workflow patterns: chaining, parallelization, routing, orchestration, plus a 'MAKER' k-voting pattern for reducing errors
• First-class MCP support, including OAuth 2.1, sampling and elicitations, so agents can call MCP servers as tools
• "Agents as tools" for decomposing larger tasks
• Multimodal support (vision, PDF, structured outputs, streaming)
• An interactive terminal shell with completions and menus, plus keyring-based secrets management

Works with

Model providers: Anthropic, OpenAI, Google, Azure, Ollama, DeepSeek, and others. Any MCP-compliant server can be plugged in as a tool source, and the framework itself is usable from any environment that can run Python.

How to install or connect

uv pip install fast-agent-mcp
fast-agent go

Minimal example from the README:

@fast.agent(instruction="Estimate object size")
async def main():
    async with fast.run() as agent:
        await agent.interactive()

Maintenance and safety

Apache-2.0 licensed, with roughly 3.9k stars, 446 forks and over 2,000 commits — an actively maintained, well-regarded project in the MCP ecosystem. No specific safety red flags found; as with any agent framework, users are responsible for scoping API credentials and vetting any MCP servers they connect.

Who should use it

Python developers who want to build custom multi-agent workflows or MCP-integrated agents from code rather than use a fixed CLI tool — good for evaluation harnesses, orchestration pipelines, and teams already invested in MCP.