What is Phidata?
Phidata is an open-source Python framework for building AI agents. It differs from LangChain and similar tools by taking a first-principles, code-first approach — agents are plain Python objects with memory, knowledge, and tools as first-class attributes, not tangled chains of callbacks.
The project has been rebranded as Agno but remains widely known as Phidata. It emphasises production readiness: agents can store long-term memory in databases, retrieve information from structured knowledge bases, and use tools like web search, SQL, Python, and REST APIs — all in a clean, readable API.
Key features of Phidata
Memory — short and long term
Agents have built-in conversation memory (session context) and long-term memory (persistent storage in PostgreSQL or SQLite) — so they remember preferences and past interactions across sessions.
Knowledge bases
Attach a searchable knowledge base — PDFs, URLs, text files, databases — to an agent. Phidata handles chunking, embedding, vector storage, and retrieval automatically.
Built-in tools
Phidata ships with ready-to-use tools: DuckDuckGo search, Python REPL, SQL queries, file operations, Exa web search, YFinance for financial data, and dozens more — attach them with a single line.
Multi-agent teams
Compose agents into teams — a lead agent can delegate sub-tasks to specialist agents (researcher, writer, coder) and combine their outputs, using any coordination strategy you define.
Any LLM backend
Works with OpenAI, Anthropic, Google Gemini, Mistral, Groq, Ollama (local models), and AWS Bedrock — swap models with one parameter change.
Agno Playground
Serve your agents locally and interact with them through the hosted Agno Playground UI — useful for testing and debugging without building a custom frontend.
How Phidata works
A basic Phidata agent is just a few lines of Python:
from phi.agent import Agent
from phi.tools.duckduckgo import DuckDuckGo
agent = Agent(tools=[DuckDuckGo()], markdown=True)
agent.print_response("What happened in AI this week?")Add memory, a knowledge base, or swap the model with additional parameters. The framework handles the tool-calling loop, context window management, and streaming internally.
Real-life use cases of Phidata
Research agents
An agent that searches the web, reads pages, and writes a structured report — all autonomously.
Data analysis agents
Connect a database and let an agent answer natural-language questions by writing and executing SQL.
Personalised assistants
An assistant that remembers your preferences, past conversations, and personal context across sessions.
Pros and cons of Phidata
Pros
- + Clean, readable Python API — minimal boilerplate
- + First-class memory and knowledge base support
- + Rich built-in tool library
- + Works with any major LLM
- + Multi-agent teams with simple composition
Cons
- − Python-only — no JavaScript SDK
- − Smaller community than LangChain
- − No visual editor — code-only workflow
- − Rebranding to Agno has caused documentation fragmentation
Phidata pricing
| Plan | Price | Details |
|---|---|---|
| Open-source | Free | Full framework on GitHub under MPL licence |
| Agno Cloud | See agno.com | Hosted playground, monitoring, and managed agent infrastructure |
Alternatives to Phidata
Tips for using Phidata
Package renamed from phidata to agno
Import paths changed from from phi.agent import Agent to from agno.agent import Agent. For new projects, install agno.
Use PostgreSQL for persistent memory in production
The default memory backend does not persist between restarts. Configure a PostgreSQL storage backend explicitly (storage=PgAgentStorage(...)) for agents that need to remember users across sessions.
Tool function docstrings are the tool description the LLM sees
The function's docstring becomes the description sent to the model. Missing type annotations or vague docstrings cause the LLM to misuse or skip the tool entirely. Write clear, specific docstrings.
Use async API in web frameworks to avoid blocking
When integrating with async web frameworks like FastAPI, always use the async API — calling the sync agent.run() from an async handler blocks the event loop under load.
Building a web research team in minutes
from phi.agent import Agent
from phi.tools.duckduckgo import DuckDuckGo
from phi.tools.newspaper4k import Newspaper4k
researcher = Agent(name="Researcher", tools=[DuckDuckGo(), Newspaper4k()])
writer = Agent(name="Writer", description="Writes clear summaries")
team = Agent(team=[researcher, writer])
team.print_response("Summarise this week's AI news")Phidata hits the sweet spot between power and simplicity — agents with real memory and tools in a handful of Python lines, without the ceremony of larger frameworks. Compare with Dify for a visual no-code approach and OpenAI Swarm for a minimal educational multi-agent framework.