DeerFlow is an open-source agent harness that orchestrates sub-agents, memory, and sandboxes to perform multi-step tasks ranging from research to code generation. It supports local or Docker-based sandbox execution and integrates with external MCP servers and messaging platforms.
Project overview
It orchestrates sub-agents, long-term memory, and sandboxed code execution to address complex tasks that can take minutes to hours.
Project type
AI Agent
Use cases
Search & Research
Deployment
Refer to project documentation
License
MIT
Best for
Developers and AI engineers needing a harness to coordinate sub-agents, memory, and sandboxed code execution for long-horizon tasks.
Enterprise teams looking to trigger automated workflows and interact with agents through IM channels such as Slack, DingTalk, or Feishu.
Key capabilities
Orchestrates sub-agents, memory, and sandboxes to perform complex long-horizon tasks.
Supports Local Execution, Docker Execution, and Docker Execution with Kubernetes for sandbox code.
Supports configurable MCP servers and skills to extend its capabilities, including HTTP/SSE and stdio MCP servers.
Supports receiving tasks from messaging apps including Telegram, Slack, Feishu, WeChat, WeCom, and DingTalk.
Ships a skill that lets users interact with a running DeerFlow instance directly from Claude Code.
Maintains long-term memory for the agent.
Includes built-in integrations for LangSmith, Langfuse, and Monocle tracing.
Limitations and risks
Deploying in untrusted environments may introduce security risks such as unauthorized illegal invocation.
Unauthorized illegal invocation can trigger bulk unauthorized requests and execute high-risk operations.
If illegally invoked for cyberattacks or data theft, the system poses compliance and legal risks.
Getting started
Setup is considered medium difficulty. It requires Python 3.12+ and Node.js 22+. The first success path involves cloning the repository, running make setup, running make install to configure the LLM provider and API keys, and starting the development environment with make dev. A starting evaluation hardware profile is 4 vCPU, 8 GB RAM, and 20 GB free SSD.
Alternatives and comparisons
An autonomous work engine that plans, acts, recovers, and proves the result by operating files, terminal, browser, and APIs while keeping the user in control.
Orchestrates swarms of CLI-based coding agents in parallel across isolated git worktrees, providing built-in terminal, review, and IDE handoff workflows.
An open-source operations platform enabling engineers to rapidly build AI Agents, Apps, and Automations that take action across business systems, which can be securely self-hosted.