DeerFlow
DeerFlow 2 is a long-horizon agent harness using subagents, memory, skills, and sandboxes. It is broader than the original research framework, which remains on a separate 1.x branch.
Utility: Research; kind: agent.
Stars: 84k+; Starred in September: 2k+.
Metrics updated: 31 Oct 2026. Profile reviewed: 31 Oct 2026.
Utilities
- Break a research or creation task into tool-using subagent work.
- Run longer tasks with sandboxed execution and persistent context.
Who it is for
- Technical teams experimenting with research and multi-step agent workflows.
Requirements
- Deploy the documented stack; Docker is a recommended installation route.
- Configure supported model credentials and the search/fetch tools you intend to use.
Limitations
- Version 2 is a ground-up rewrite, so version-1 instructions are not interchangeable.
- The README requires network isolation for self-hosted fetch backends; do not casually expose private networks.
Cost and privacy notes
- Multiple subagents, model calls, and search tools can compound usage costs.
- Review sandbox storage, provider routing, and outbound-network isolation before using internal documents.
Illustrative example (untested)
Illustrative, not tested: request a source-linked comparison of two developer tools and review each supporting reference.
License
MIT; reviewed the repository LICENSE.