What it does

PraisonAI is a framework for building agents, from a single agent to teams that hand work to one another. Its README shows an agent created in a few lines of Python from a set of instructions and started with a single call. A JavaScript package is available through npm, and a command-line tool installs with pip. The README organises the framework into five layers: prompt, context, harness, loop and graph. Context covers memory, knowledge lookup and automatic compaction. The harness covers tools, MCP connections, guardrails and approval gates. The loop sets iteration and budget limits and catches agents that repeat themselves. The graph layer chains, routes and parallelises agents, in Python or in YAML. Agent tools can run in a Docker container or in hosted sandboxes such as E2B, Modal, Daytona and Fly.io. The README lists 24 model providers, including OpenAI, Anthropic, Google Gemini, Ollama and Groq. Optional add-ons include a dashboard that connects agents to Telegram, Slack or Discord, a drag-and-drop flow builder and a chat interface. The project is open source under the MIT licence. The quick start sets an OpenAI API key, so model usage is billed by whichever provider you configure.

Capabilities

Single agents or multi-agent teams with handoffs, routing and parallel steps
Memory, knowledge lookup and context compaction
MCP tool connections, guardrails and human approval gates
Budget limits and detection of agents stuck in loops
Tools can run in Docker or hosted sandboxes
Works with 100+ language models, including local ones through Ollama

When to use it

You want to build agent systems in Python or JavaScript instead of using a hosted app
You need agents that keep memory, call tools and pass work to each other
You want to connect agents to Telegram, Slack or Discord
You want to run the same agents on different model providers