OpenManus

The core is a single Manus agent built on a ReAct-style tool-call loop. Each step the model picks tools (local ones such as PythonExecute and StrReplaceEditor, plus tools loaded from connected MCP servers), the results are appended to the conversation, and the loop ends when the agent calls terminate or reaches max_steps. A duplicate-response check nudges the agent when it repeats itself.

The multi-agent path is PlanningFlow, started by run_flow.py. The flow's LLM uses a planning tool to write a step list, labelling steps with agent names; the flow then walks the plan, hands each step to the agent whose key matches the label (or the primary agent), and marks step status until the plan is done. The README calls this multi-agent version unstable.

State is in-process only. Each agent keeps a Memory object holding the message list, trimmed to the most recent 100 messages; there is no documented persistence, vector memory or resume from checkpoint.

The default Manus agent includes an ask_human tool; when the model calls it, the process prints the question and waits for a typed answer in the terminal (input()). A run can be stopped with Ctrl+C, which the entry scripts catch, and run_flow.py aborts after a 60-minute timeout. No approval step before Python, bash or file-edit tools run is documented.

git clone https://github.com/FoundationAgents/OpenManus.git && cd OpenManus && uv venv --python 3.12 && source .venv/bin/activate && uv pip install -r requirements.txt

# Run from the OpenManus repo root after copying config/config.example.toml
# to config/config.toml and adding your model key there.
import asyncio

from app.agent.manus import Manus
from app.flow.flow_factory import FlowFactory, FlowType


async def main():
    flow = FlowFactory.create_flow(flow_type=FlowType.PLANNING, agents={"manus": Manus()})
    result = await flow.execute("Draft a README for a small to-do CLI and save it as workspace/README.md")
    print(result)


asyncio.run(main())

• The README creates the environment with Python 3.12 (conda or uv venv --python 3.12).
• Nothing runs until config/config.toml exists: copy config/config.example.toml and add an API key and model.
• Browser control uses uvx browser-use --cli-mcp, so uv/uvx must be available; set OPENMANUS_DISABLE_BROWSER_USE=1 to turn it off. BrowserGym still needs playwright install.
• The multi-agent run_flow.py is labelled unstable in the README; the DataAnalysis agent needs extra dependencies from app/tool/chart_visualization/README.md.
• python main.py runs the single agent; python run_mcp.py runs the MCP tool version.

Yes. The agent connects to MCP servers (Browser Use CLI runs as the default one, and more can be added in a JSON config), and run_mcp_server.py exposes OpenManus tools as an MCP server.

Partly. The repo contains an experimental A2A server for the Manus agent that only supports non-streaming mode and pins a2a-sdk 0.2.5. AG-UI support was not found.

Yes, it is MIT licensed with no paid tier. You pay for the model API you configure; optional Browser Use Cloud and Daytona sandboxes need their own API keys.

Run python run_flow.py, which plans the task and hands steps to agents; set use_data_analysis_agent = true under [runflow] to add the DataAnalysis agent. The README marks this mode as unstable.

It is developed under the FoundationAgents GitHub organization by contributors from MetaGPT, according to the README.