Summary

DeerFlow 2.0 is an open-source super agent harness designed to orchestrate sub-agents, memory, and sandboxes for executing complex tasks. It targets developers and power users seeking an extensible AI runtime for automation, offering a robust infrastructure for agents to perform multi-step operations efficiently.

How It Works

This ground-up rewrite leverages LangGraph and LangChain to function as a "super agent harness." It orchestrates sub-agents, memory, and sandboxes, powered by extensible "skills" (structured capability modules). This approach enables agents to decompose complex tasks, spawn sub-agents for parallel execution, and utilize a sandboxed environment with a full filesystem. Skills are loaded progressively, maintaining a lean context window for token efficiency.

Quick Start & Requirements

• Primary Install/Run: Docker is recommended (make docker-init, make docker-start). Local development requires cloning the repo, running make config, and then make dev.
• Prerequisites: Local development requires Node.js 22+, pnpm, uv, and nginx. API keys for configured LLM providers (e.g., OpenAI, OpenRouter, Claude Code) are essential.
• Setup: Configuration involves running make config, editing config.yaml to define models, and setting API keys via a .env file or environment variables.
• Links: Official Website (URL not provided), Documentation, Contributing Guide.

Highlighted Details

• Extensible Skills & Tools: Ships with built-in skills (research, report generation, web pages) and core tools (web search, file operations). Supports custom skills via MCP servers and Python functions, loaded progressively.
• Sandboxed Execution: Tasks run within isolated Docker containers featuring a dedicated filesystem for skills, workspace, uploads, and outputs. Supports local, Docker, and Kubernetes sandbox modes.
• Sub-Agent Orchestration: The lead agent can spawn sub-agents dynamically for parallel execution and complex task decomposition, each with scoped contexts and termination conditions.
• Claude Code Integration: A claude-to-deerflow skill enables direct interaction with DeerFlow from Claude Code for task management and status checks.
• Long-Term Memory: Builds a persistent, locally stored memory across sessions for user profiles, preferences, and accumulated knowledge.

Maintenance & Community

Core authors include Daniel Walnut and Henry Li. The project acknowledges community contributions but does not list specific community channels (e.g., Discord/Slack) or sponsorships in the README.

Licensing & Compatibility

Licensed under the MIT License. Permissive for commercial use and integration with closed-source projects.

Limitations & Caveats

DeerFlow 2.0 is a complete rewrite of v1, sharing no code. Configuration requires careful setup of LLM API keys and model definitions. Some CLI-backed providers may have specific authentication or token limitations.