DeerFlow vs OpenHands
DeerFlow 2.0 is ByteDance's self-hosted agent harness built on LangGraph: a lead agent with sandboxes, skills, memory and sub-agents, aimed at research, reports and coding. OpenHands centers on coding: Agent Canvas (beta) manages the OpenHands agent or Claude Code, Codex and Gemini CLI across local, Docker, VM or cloud backends. Both are MIT-licensed.
Facts side by side
| Fact | DeerFlow | OpenHands |
|---|---|---|
| Type | Harness | Harness |
| Languages / SDKs | Python, TypeScript | TypeScript, Python |
| License | MIT | MIT |
| Pricing model | Open source, free | Open core |
| Orchestration pattern | Supervisor | Supervisor |
| GitHub stars | 83,261 (as of 2026-09-30) | 89,602 (as of 2026-09-30) |
| GitHub forks | 11,552 | 11,827 |
| Last push | 2026-09-30 | 2026-09-30 |
| Latest release | v2.1.0 | v1.24.0 |
| Repository | bytedance/deer-flow | OpenHands/OpenHands |
| Website | deerflow.tech | openhands.dev |
| Documentation | github.com | docs.openhands.dev |
| Last verified | 2026-09-30 | 2026-09-30 |
| MCP support | Yes (checked 2026-09-30) | Yes (checked 2026-09-30) |
| A2A support | Unknown (checked 2026-09-30) | Unknown (checked 2026-09-30) |
| AG-UI support | Unknown (checked 2026-09-30) | Unknown (checked 2026-09-30) |
| Install verified | Not yet | Not yet |
Choose DeerFlow if
- Your work is mostly research and long reports, with coding as one of several task types.
- You want durable batch work:
batch_taskruns large independent item sets from a SQL-backed queue that survives restarts. - You want tasks to arrive from chat channels (Telegram, Slack, Discord, Feishu/Lark, DingTalk, WeChat, WeCom) or to embed the backend in Python with
DeerFlowClient.
Choose OpenHands if
- Your main workload is software engineering, run as always-on coding agents and automations.
- You want per-conversation isolation (one container per conversation) and a choice of host, Docker, VM, Kubernetes or managed cloud backends.
- You want an npm install and a hosted option: OpenHands is open-core with a paid cloud, while DeerFlow documents a git clone plus make targets and no hosted service.
Migration notes
The two share a pattern (a supervising agent that spawns children and can call Claude Code or Codex over ACP) but no code or state format. DeerFlow keeps state in LangGraph checkpoints and a store with DeerMem long-term memory; OpenHands' Agent Server persists conversations that can be resumed or branched. Sub-agent definitions do not transfer: DeerFlow configures sub-agents per custom agent, OpenHands defines them as Markdown files with frontmatter. Human review differs: DeerFlow decides per run whether a human may be asked (scheduled and webhook runs never wait), OpenHands uses confirmation policies (AlwaysConfirm, NeverConfirm, ConfirmRisky). Both warn about host access: DeerFlow is designed for a trusted local environment, and OpenHands' quick npm install runs its agent server with full filesystem access unless Docker, a VM or a cloud backend is used. DeerFlow 2.x shares no code with 1.x.
FAQ
Is DeerFlow a coding agent?
Coding is one of its task types. The project describes 2.0 as a super agent harness for research, coding and content work, built on LangGraph, with sub-agents and sandboxes.
Can both run Claude Code and Codex?
Yes. DeerFlow has an invoke_acp_agent tool for ACP agents, and OpenHands can run Claude Code, Codex or Gemini CLI as ACP agents inside a conversation.
Which one has a hosted version?
OpenHands has a paid cloud (pricing page linked on its tool page). No hosted DeerFlow service is documented.
What hardware does DeerFlow need?
The README suggests at least 4 vCPU and 8 GB RAM for local evaluation.