How foundationagents/openmanus works
OpenManus is a tool-using LLM agent platform. In the principal workflow, a user submits a task through the terminal entry point; Manus initializes MCP integrations, asks the LLM for tool-directed decisions, dispatches tools, and returns an outcome. The repository also includes multi-agent planning, data-analysis and browser agents, sandbox-backed tools, an MCP server, and an A2A adapter. The graph combines shown calls with documented capabilities and source-index references; some unsampled subsystem links are omitted or marked unverified.
Main entry points: manus.py Manus agent · toolcall.py Tool-call loop · llm.py LLM client · base.py Tool interface
Components 21
Agent runtime
- Terminal entry
main.py - Manus agent
app/agent/manus.py - Tool-call loop
app/agent/toolcall.py - LLM client
app/llm.py - Planning flow
app/flow/planning.py - Browser agent
app/agent/browser.py - Data-analysis agent
app/agent/data_analysis.py - App configuration
app/config.py
Tools and capabilities
- Tool interface
app/tool/base.py - File editor
app/tool/str_replace_editor.py - Web search
app/tool/web_search.py - Data visualization
app/tool/chart_visualization/data_visualization.py - MCP tools
app/tool/mcp.py
Sandbox execution
- Sandbox tools
app/tool/sandbox - Daytona tool base
app/daytona/tool_base.py - Daytona lifecycle
app/daytona/sandbox.py - Sandbox runtime
app/sandbox/core/sandbox.py
Protocols and integrations
- MCP server
app/mcp/server.py - MCP services
- A2A adapter
protocol/a2a/app/agent.py
Other components
- Task submitter
Connections 20
- Task submitter → Terminal entry: submits task
README.md - Terminal entry → Manus agent: creates agent
main.py - Manus agent → Tool-call loop: extends
app/agent/manus.py - Tool-call loop → LLM client: requests decisions
app/agent/toolcall.py - Tool-call loop → Tool interface: dispatches tools
- Manus agent → MCP services: connects tools
app/agent/manus.py - Manus agent → App configuration: loads settings
app/agent/manus.py - Planning flow → LLM client: uses model
app/flow/planning.py - Planning flow → Tool interface: uses tools
app/flow/planning.py - Browser agent → Tool-call loop: uses loop
- Data-analysis agent → Data visualization: supports analysis
- Data visualization → LLM client: uses model
app/tool/chart_visualization/data_visualization.py - Web search → MCP services: queries engines
- MCP server → Tool interface: registers tools
app/mcp/server.py - MCP server → MCP services: serves requests
- Sandbox tools → Daytona tool base: uses base
app/agent/sandbox_agent.py - Daytona tool base → Daytona lifecycle: creates sandbox
app/daytona/tool_base.py - Sandbox runtime → App configuration: reads settings
app/sandbox/core/sandbox.py - A2A adapter → Manus agent: adapts agent
protocol/a2a/app/agent.py - Manus agent → File editor: offers tool
app/agent/manus.py