DeerFlow orchestrates sub-agents, memory, and sandboxes to do almost anything — powered by extensible skills. Built on LangGraph and LangChain.
DeerFlow started as a Deep Research framework. The community ran with it — building data pipelines, generating slide decks, spinning up dashboards. So we rebuilt it from scratch.
Skills are structured capability modules — Markdown files that define workflows, best practices, and resources. Loaded progressively, only when needed.
The lead agent spawns sub-agents on the fly — each with scoped context, tools, and termination conditions. They run in parallel and report structured results.
Each task gets its own execution environment with a full filesystem view — skills, workspace, uploads, outputs. Shell execution runs inside isolated containers.
Across sessions, DeerFlow builds persistent memory of your profile, preferences, and accumulated knowledge. Memory is stored locally under your control.
DeerFlow doesn't just talk about doing things. It has its own computer.
Run the setup wizard to choose an LLM provider, web search, and execution preferences. Generates config.yaml in about 2 minutes.
Choose Docker (recommended) or local development. One command to start — sandbox, gateway, frontend, all wired up.
Send tasks via web UI, IM channels (Telegram, Slack, Feishu, WeChat), or Claude Code. Fan out into sub-agents automatically.
The lead agent synthesizes everything — reports, websites, slide decks, dashboards — into a coherent output.
DeerFlow 2.0 ships with everything an agent needs out of the box.
Built-in skills for research, report generation, slide creation, web pages, image and video generation. Extend with custom skills via MCP servers or Python functions.
Decompose complex tasks into parallel sub-agents, each with isolated context and tools. The lead agent synthesizes everything into a coherent output.
Local, Docker, or Kubernetes sandboxes. Isolated containers with full filesystem access — uploads, workspace, outputs.
Receive tasks from Telegram, Slack, Feishu, DingTalk, WeChat, and WeCom. No public IP needed — all channels auto-start.
Aggressive summarization, isolated sub-agent contexts, strict tool-call recovery. Stays sharp across long multi-step tasks.
Persistent memory across sessions. Your profile, preferences, and accumulated knowledge — stored locally, under your control.
Clone the repo, run one command, and hand the harness your first task. Docker recommended — two minutes to a running instance.
DeerFlow is an open-source super agent harness — a runtime that gives agents the infrastructure to get work done. It orchestrates sub-agents, memory, and sandboxes to handle tasks that take minutes to hours.
DeerFlow is built by ByteDance and lives at github.com/bytedance/deer-flow. It claimed the #1 spot on GitHub Trending on February 28th, 2026.
DeerFlow is model-agnostic and works with any LLM that implements the OpenAI-compatible API. It performs best with models supporting long context windows (100k+ tokens) and strong tool-use capabilities.
DeerFlow is not a framework you wire together — it's a harness. Batteries included: filesystem, memory, skills, sandbox-aware execution, and the ability to plan and spawn sub-agents for complex multi-step tasks.
Docker is the recommended deployment target. Run make up for production or make docker-start for development. The unified endpoint runs on port 2026.
DeerFlow is designed for local trusted environments. For production, use IP allowlists, authentication gateways, and network isolation. See the Security Notice in the repo for detailed recommendations.