Agno: from Phidata to a self-hosted Python agent platform

From agent framework to “agent platform”

Agno started life as Phidata and has since repositioned itself. Rather than just helping you write an AI agent, it aims to cover the whole stack. The Python SDK defines agents, teams that form a multi-agent system, and workflows. The AgentOS runtime then serves them as a REST API with streaming, stores sessions, memory, knowledge and traces in your own database, and handles authentication with role-based access control. A web management UI sits on top.

You also get ready-made toolkits (GitHub, Slack, Postgres and more), human approval before sensitive actions, OpenTelemetry-based observability, a built-in scheduler and interfaces for Slack, Telegram, WhatsApp and Discord. Agents can be exposed as an MCP server or through the A2A protocol.

Getting started

The repository offers two paths. The fastest is to hand a coding assistant a short prompt that clones a deployment template (Railway, Docker, AWS, Google Cloud, Azure and others) and runs the platform locally with Docker: you end up with an API, a Postgres database, an MCP server and a control plane. If you would rather write the code yourself, the docs walk you through a first agent in a few lines, then through building your own platform. You will need an API key from a model provider.

Strengths and limits

Strengths. Apache 2.0 license, an “own your stack” philosophy with data and traces kept on your side, production features built in (API, storage, security, scheduling), and broad model provider support.

Limits. The platform ambition makes it heavier than a plain library: Docker and a database are part of the recommended path. The rename and fast-moving APIs mean that many older tutorials, written for Phidata, are out of date. Telemetry is on by default. And connecting the control plane to a live production deployment is a paid feature.

Alternatives

In Python, CrewAI offers a role-based multi-agent approach that is easier to pick up, LangGraph gives very fine control over execution graphs, and Pydantic AI is a lighter library. On the TypeScript side, Mastra follows a similar all-in-one framework logic. The right pick mostly depends on how much infrastructure you want to run yourself.

Official site: agno.com.