DeerFlow is ByteDance's open-source "long-horizon SuperAgent harness" — a ground-up rewrite of their Deep Research framework that orchestrates parallel sub-agents, persistent cross-session memory, sandboxed execution, and extensible skills to handle tasks that take minutes to hours. Built on LangGraph/LangChain with MIT licence.
What It Is
Where tools like Claude Code and OpenHarness focus on coding tasks in a developer's terminal, DeerFlow targets a different problem: multi-step, multi-hour workflows — researching a topic across dozens of sources, writing a report, generating a slide deck, automating a data pipeline. A lead agent decomposes the task, spawns parallel sub-agents with isolated contexts, and synthesises their results into structured output.
The v2 rewrite (February 2026) replaced v1's sequential deep-research architecture with a proper multi-agent orchestration layer, adding sandboxed execution, persistent memory, and IM channel integration (Telegram, Slack, Feishu, WeChat, WeCom).
Architecture
| Layer | Implementation |
|---|---|
| Agent runtime | LangGraph orchestration — lead agent spawns and coordinates sub-agents |
| Gateway API | REST interface for task submission, model management, skill configuration |
| Sandbox | Local, Docker, or Kubernetes — isolated filesystem access per task |
| Skills | On-demand modular capabilities (research, reporting, web operations, slide creation) |
| Memory | Cross-session persistent knowledge; session-aware context compression |
| IM gateway | Native Telegram, Slack, Feishu, WeChat, WeCom integration |
| Observability | LangSmith and Langfuse tracing |
The multi-service architecture (Python backend + Node.js/TypeScript frontend + optional Kubernetes provisioner) makes DeerFlow significantly heavier than single-process harnesses like OpenHarness or Goose.
How It Differs from Claude Code / OpenHarness
| DeerFlow | Claude Code | OpenHarness | |
|---|---|---|---|
| Primary use case | Long-horizon research & automation | Coding tasks in terminal | Coding tasks (any provider) |
| Agent model | Lead + parallel sub-agents | Single agent + subagents | Single agent + subagents |
| Memory | Cross-session persistent | Session only (MEMORY.md) | Session + MEMORY.md |
| Sandbox | Docker / Kubernetes | Host environment | Host environment |
| Built on | LangGraph | Custom | Custom |
| IM integration | Telegram, Slack, Feishu | None | Slack, Feishu (via ohmo) |
| Deployment | Multi-service (frontend + API) | CLI | CLI |
| License | MIT | Proprietary | MIT |
Claude Code Integration
DeerFlow ships a /claude-to-deerflow skill that lets a running Claude Code session submit tasks to a local DeerFlow instance — useful for offloading long-horizon research subtasks while continuing coding work in Claude Code.
Why Assess
ByteDance backing and 61,000+ GitHub stars after the v2 launch (ranked #1 GitHub Trending on February 28, 2026) are strong signals. But:
- Heavyweight deployment: Production sizing guide calls for 16 vCPU, 32 GB RAM, 40 GB SSD — substantially more than a terminal agent. Not a drop-in tool for individual developers.
- LangGraph dependency: Inherits LangGraph's complexity and API churn risk. Teams not already in the LangGraph ecosystem face a steeper setup cost.
- Use-case specificity: The long-horizon research-and-automation niche has less immediate overlap with most engineering teams' day-to-day AI usage than a coding agent like Claude Code or Goose.
- No named external adopters: ByteDance presumably uses this internally, but no external organisations are cited.
Assess means: experiment with DeerFlow for multi-step research or content automation workflows where a coding agent is the wrong tool. Track whether the lightweight "Gateway mode" (experimental single-runtime without a separate LangGraph server) reduces the operational burden.
Getting Started
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make setup # interactive configuration wizard
make docker-start # recommended for first run
# Access: http://localhost:2026
Key Characteristics
| Property | Value |
|---|---|
| Interface | Web UI + REST API |
| Provider | ByteDance |
| License | MIT |
| Language | Python (68.7%), TypeScript (19.4%) |
| Built on | LangGraph + LangChain |
| GitHub stars | ~61,100 (April 2026) |
| GitHub | bytedance/deer-flow |
| Website | deerflow.tech |
Sources
- bytedance/deer-flow — GitHub — official repository; README, architecture docs
- DeerFlow — deerflow.tech — project site
- Deer Flow: ByteDance's Open-Source Long-Horizon SuperAgent Framework (pyshine.com) — practitioner overview