lionagi 发布多智能体编排框架，支持 CLI 与 Python 两种用法

lionagi is a governed multi-agent orchestration framework. Build agent workflows in Python with typed, inspectable state, or run them straight from the terminal with the li CLI: single agents, parallel fan-outs, and DAG flows where an orchestrator plans specialist workers. Built continuously since 2023.

Docs | How lionagi Compares | Discord | PyPI | Changelog

Demo

Lion Studio walkthrough (13 min): inspecting runs, checking agent claims against actual output, and launching work with approval gates. Built on lionagi, runs locally.

Quick start

pip install lionagi

# one agent
li agent claude/sonnet "Explain the observer pattern in 3 sentences"

# three workers in parallel, then a synthesis pass
li o fanout claude/sonnet "Identify code smells in this codebase" -n 3 --with-synthesis

# an orchestrator plans a DAG of specialists; workers run as dependencies resolve
li o flow claude/sonnet "Audit the auth module for security issues" --cwd .

CLI model aliases (claude, codex, ...) spawn the provider's own CLI as a subprocess, so an existing claude login subscription works with no API key. API providers take the usual environment keys (OPENAI_API_KEY, ...). Details in Providers.

The same engine from Python, with structured output as a first-class result:

import asyncio
from pydantic import BaseModel
from lionagi import Branch

class Assessment(BaseModel):
    risk: str
    reasons: list[str]

async def main():
    b = Branch(chat_model="codex/gpt-5.5", system="You are a careful reviewer.")
    result = await b.operate(
        instruction="Assess the risk of enabling auto-merge on this repository.",
        response_format=Assessment,
    )
    print(result.risk, result.reasons)

asyncio.run(main())

Why lionagi

• You own the loop. Branches, sessions, and flows are ordinary Python objects and CLI commands. There is no framework runtime to surrender control to, and no hidden prompt assembly between you and the model.
• Typed, inspectable state. Every conversation is a collection of typed messages with an explicit ordering. State serializes, persists, and resumes; it is never an opaque blob inside a chain.
• CLI agents are first-class endpoints. Claude Code, Codex, and other coding CLIs sit in the same orchestration graph as API models, so subscription-based agents and API calls compose in one flow.
• Durable runs. Every run persists under ~/.lionagi/runs/. Resume any branch with li agent -r, reattach with -c, watch live work with li monitor, and schedule recurring runs with li schedule.
• Governance built in. Permission policies per tool call, guard hooks that block destructive commands and off-limits paths, and git-worktree sandboxing for speculative edits that never touch your branch until you merge them.

For the architecture-level comparison with LangChain / LangGraph (and a field matrix covering LlamaIndex and AG2), see How lionagi Compares.

Concepts

The li CLI

# Single agent, resumable
li agent claude/sonnet "Explain the observer pattern in 3 sentences"
li agent -r <branch-id> "follow up on your findings"

# Fan-out: N workers in parallel, optional synthesis
li o fanout claude/sonnet "Identify code smells in this codebase" -n 3 --with-synthesis

# DAG flow: orchestrator plans agents with dependency edges
li o flow claude/sonnet "Audit the auth module for security issues" --cwd .

# Playbook: parametric flow spec at ~/.lionagi/playbooks/audit.playbook.yaml
li play audit --mode security "the auth service"

# Team messaging: inbox coordination between agents
li team create "review" && li team send "Start analysis" -t <id> --to analyst

# Observe and operate
li monitor --since 1h              # live and recent sessions, flows, plays
li schedule create ...             # cron / interval / repo-event triggers
li kill <id>                       # stop a running session or invocation

# Time-bounded run: injects a [DEADLINE] preamble so the agent paces itself
li agent claude/sonnet --timeout 300 "Audit the auth module and produce a summary"

Every command and flag: CLI Reference. Installable playbook and skill templates: examples/.

Python API

Chat

from lionagi import Branch

b = Branch(chat_model="openai/gpt-5.4", system="You are a concise assistant.")
reply = await b.communicate("What causes rainbows?")

Structured output

from pydantic import BaseModel

class Summary(BaseModel):
    points: list[str]
    confidence: float

result = await b.operate(instruction="Summarize this text.", response_format=Summary)

Tools + ReAct

from lionagi.tools.types import ReaderTool

branch = Branch(tools=[ReaderTool])
result = await branch.ReAct(
    instruct={"instruction": "Summarize /path/to/paper.pdf"},
)

Full reference → docs/api/

Lion Studio

Lion Studio is the built-in web UI for operating your agent workflows: projects, schedules, playbooks, execution DAGs, and run inspection in one place.

No download or build required. Lion Studio is a client-side web app — open lion-studio.khive.ai directly in your browser and it connects to your local daemon at http://127.0.0.1:8765. Your data never leaves your machine; the hosted page is just the UI. Available in 16 languages: English, 中文, Español, Français, हिन्दी, বাংলা, Deutsch, Bahasa Indonesia, Português (Brasil), 한국어, Türkçe, اردو, Tiếng Việt, العربية, Русский, and 日本語 — with full right-to-left support.

li studio             # default: starts the local daemon and opens
                      # https://lion-studio.khive.ai — nothing is built locally
li studio --docker    # self-contained: auto-pulls ghcr.io/ohdearquant/lion-studio
                      # UI + API → http://localhost:8765
li studio --no-frontend  # backend API only, no UI
li studio --dev       # from a source checkout: backend + in-repo frontend, hot reload

--web, --docker, --no-frontend, and --dev are mutually exclusive; --web is the default when no flag is given. Pass --no-open to skip auto-opening the hosted UI in your browser.

Providers

CLI aliases spawn subprocess tools, not REST API calls:

• claude: install Claude Code CLI → claude login (subscription) or export ANTHROPIC_API_KEY=sk-ant-...
• codex: requires ChatGPT Plus/Pro → npm install -g @openai/codex → codex login
• deepseek: export DEEPSEEK_API_KEY=sk-...
• pi: install Pi Code CLI
• Python API (iModel, Branch): export OPENAI_API_KEY=sk-... for the default model

API-endpoint providers (OpenAI, Anthropic, Gemini, Ollama, NVIDIA NIM, Perplexity, Groq, OpenRouter): Providers reference.

Docs

Optional Extras

uv add "lionagi[reader]"    # Document reading (PDF, HTML, DOCX)
uv add "lionagi[mcp]"       # MCP server support
uv add "lionagi[ollama]"    # Local models via Ollama
uv add "lionagi[rich]"      # Rich terminal output
uv add "lionagi[graph]"     # Flow visualization
uv add "lionagi[postgres]"  # PostgreSQL persistence
uv add "lionagi[all]"       # Everything

Claude Code Marketplace

Installable Claude Code plugins for the lionagi agent runtime — structured show runs, memory management, playbook authoring, developer tooling, and multi-agent orchestration:

claude /plugin marketplace add ohdearquant/lionagi

Full plugin list: marketplace/README.md.

Community

• Discord — questions, ideas, help
• Issues — bugs and feature requests
• Contributing — PR workflow

Citation

@software{Li_LionAGI_2023,
  author = {Haiyang Li},
  year   = {2023},
  title  = {LionAGI: Towards Automated General Intelligence},
  url    = {https://github.com/ohdearquant/lionagi},
}