DeerFlow 🤖 Agent Open source
ByteDance's self-hosted agent harness with subagents, sandbox, memory and skills
- GitHub stars
- 83k
- Stars this week
- +360
- Forks
- 12k
- Licence
- MIT
- Last push
- 2026-10-02
- Maintainer
- bytedance
git clone https://github.com/bytedance/deer-flow.gitThird-party subagents & agents run with your permissions. Read the source before installing, and prefer pinned versions.
Works with
About DeerFlow
What it does
DeerFlow is ByteDance's open-source super agent harness. Version 2 is a full rewrite built on LangGraph and LangChain: a lead agent plans work, spawns subagents, runs code in a sandbox, keeps long-term memory and loads skills on demand. It started as a deep-research tool and now also writes code, builds web pages and produces reports and slides. It is a self-hosted application with a web workspace, not a plugin for an existing editor.
What is inside
The stack includes a gateway, a web UI, a terminal workbench and an embedded Python client. Skills follow the standard SKILL.md format and load progressively, and you can import .skill archives or export your own. MCP servers can be added with OAuth support, and the sandbox can run locally or in containers. Other parts cover session goals, manual context compaction, projects with document shelves, scheduled tasks, browser control, and chat connectors for Slack, Telegram, Feishu and others. Tracing hooks exist for LangSmith, Langfuse and Monocle. A claude-to-deerflow skill lets Claude Code send tasks to a running DeerFlow instance and manage its threads.
Works with
DeerFlow runs on its own. It can use Codex CLI or Claude Code credentials as model providers, and the bundled skill connects Claude Code to it.
How to install
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make setup
npx skills add https://github.com/bytedance/deer-flow --skill claude-to-deerflow
Maintenance and safety
MIT licensed, very active and backed by a large team. The README carries a clear security notice: the agent can run system commands, it binds to localhost by default, and gateway admin access equals code execution on the host. Do not expose it without an allowlist and authentication in front. Setup needs Docker or Python 3.12 and Node 22.
Who should use it
Teams that want a self-hosted, general agent for research-plus-code tasks and are comfortable running a multi-service stack. If you just want help inside your editor, a coding agent plugin is a lighter choice.
Pros
- Backed by a large team with very active development
- Standard SKILL.md skills, MCP servers and sandboxed execution
- Clear security guidance for deployment
Cons
- Heavy multi-service stack aimed at general research tasks, not an editor plugin
- Gateway admin equals code execution, so it must stay locked down
Similar subagents & agents
All agent workflows & frameworks →Spec Kit 🤖 AgentFree
GitHub's toolkit for spec-driven development with AI coding agents
Task Master 🤖 AgentFree
AI task management that turns a PRD into tasks your coding agent works through
Fast Agent 🤖 AgentOpen source
Python framework for building, orchestrating and evaluating MCP-native AI agents
OpenSpec 🤖 AgentFree
Lightweight spec-driven development: agree on changes before the agent codes
CC Safety Net 🤖 AgentOpen source
Pre-execution guard that blocks destructive commands and secret access for coding agents
BMAD Method 🤖 AgentFree
Agile AI-driven development with analyst, PM, architect, developer and UX agents