PraisonAI

The core unit is an Agent (instructions or role/goal, an LLM, tools). For teams, AgentTeam takes agents plus Task objects: with process="sequential" (the default) tasks run in order and each receives the previous output; with process="hierarchical" a manager_llm decides which agent handles which task and gathers the results, retrying manager parse failures up to three times.

AgentFlow is a separate class for explicit workflows: a list of steps where route() branches on a classifier result, parallel() fans out and joins, and repeat() loops an agent until a condition holds. Agents can also hand a conversation to another agent through handoffs=[...], optionally filtering what history the receiver sees.

State and memory are configured per agent: memory={...} with a user_id keeps memory across runs, a db(...) entry persists messages, runs and traces to databases such as PostgreSQL or SQLite, and context="summarize" compacts the window before it overflows. Loop limits (max_iter, budget caps) and doom-loop detection bound each run, and tools_run_on= can move shell, file and code tools into a sandbox such as Docker or E2B.

Setting approval=True on an agent makes it pause before risky tools run; the approval docs show routing these requests to Slack, Discord, Telegram, a webhook or a local dashboard, and the CLI exposes the same choice with --approval slack and similar flags. Plan approval for planning=True is a separate mechanism, and a SQLite-backed durable approval store keeps pending bot approvals across restarts. In autonomous runs, completion_reason can come back as needs_help, handing control back to the caller.

pip install praisonaiagents

from praisonaiagents import Agent, AgentTeam, Task

# Uses OPENAI_API_KEY; pass llm="ollama/llama3.2" to an Agent for a local model.
researcher = Agent(name="Researcher", role="Researcher", goal="Collect key facts")
writer = Agent(name="Writer", role="Writer", goal="Write short, clear summaries")

research = Task(
    description="List three facts about the Rhine river.",
    expected_output="Three bullet points",
    agent=researcher,
)
summary = Task(
    description="Turn the facts into one short paragraph.",
    expected_output="One paragraph",
    agent=writer,
    context=[research],
)

team = AgentTeam(agents=[researcher, writer], tasks=[research, summary], process="sequential")
print(team.start())

• praisonaiagents requires Python 3.10 or newer.
• Package split: praisonaiagents is the SDK only, praisonai adds the CLI and extras, and praisonai-code is the CLI without gateway and bot integrations.
• The default provider is OpenAI (OPENAI_API_KEY); local Ollama models need the ollama/ prefix in llm.
• The README's multi-agent snippet imports Agents, which is now a deprecated alias of AgentTeam and emits a warning.
• process="hierarchical" needs manager_llm.
• External coding CLIs (Claude Code, Codex, Gemini CLI, Cursor CLI) must be installed and logged in separately; the Claude Code example uses skip_permissions=True, which skips that CLI's own permission prompts.

Yes, all three in its own docs: MCP() connects to MCP servers over stdio, HTTP, WebSocket or SSE; agents can be A2A servers and use A2AClient; and the AGUI class serves an agent over AG-UI for CopilotKit-style frontends.

It is MIT licensed and no paid plan or pricing page was found on the project site. Model calls are billed by whichever provider you configure.

pip install praisonaiagents gives the Python SDK only; pip install praisonai adds the CLI and optional extras; npm install praisonai installs the TypeScript SDK.

Yes. Integration classes wrap the Claude Code and Codex CLIs (plus Gemini CLI and Cursor CLI) as tools, and an Agent can use the claude CLI as its model backend. The CLIs must be installed separately.

Sequential runs tasks in list order, passing each output forward. Hierarchical adds a manager LLM (manager_llm is required) that assigns tasks to agents and collects their reports.