What is DeerFlow?
DeerFlow (Deep Exploration and Efficient Research Flow) is ByteDance's open-source agent harness, built on LangGraph and LangChain. It started as a deep-research framework; version 2.0, launched in February 2026 and tagged v2.0.0 in June, is a ground-up rewrite that shares no code with v1 (still maintained on a 1.x branch) and repositions it as a general harness that plans, spawns sub-agents, runs code in sandboxes and keeps long-term memory.
It runs as a Docker stack or local dev server with a web workspace, a Gateway API, an embedded Python client, a terminal workbench and IM channels such as Telegram, Slack, Discord and Feishu. It is model-agnostic over OpenAI-compatible APIs, with presets for OpenAI, Anthropic, DeepSeek, Qwen, Doubao, Kimi, Zhipu and MiniMax.
The repo is MIT-licensed with about 84,800 GitHub stars and roughly 990 open issues; the latest release is v2.1.0 (24 September 2026) and commits land daily.
Is DeerFlow worth it?
DeerFlow 2.0 is less a framework than a product you deploy: a server with sub-agents, a sandboxed file system, memory, scheduled tasks, projects and chat channels already wired together. That is its main advantage over LangGraph or CrewAI, where you assemble those parts yourself, and its main cost, because the surface you now operate is large. The README alone runs to over 4,000 lines of configuration notes.
The rewrite follows what users did with v1. ByteDance says people pushed the deep-research tool into data pipelines, slide decks and dashboards, so 2.0 gives the agent a workspace and lets the lead agent hand pieces of a job to sub-agents with their own context. Execution can run on the host, in Docker containers, or in Kubernetes pods through a provisioner, and the setup wizard asks up front whether to allow bash and file writes. Search and fetch plug into a long list of providers (Tavily, Brave, Jina, SearXNG, Firecrawl and others), and a claude-to-deerflow skill lets Claude Code send it jobs.
Security is the part to read twice. The project's own notice says DeerFlow can execute system commands and is designed for a trusted machine on 127.0.0.1; the Docker stack binds to loopback by default, and opening it to a LAN or the internet without an IP allowlist and an authenticating proxy is, in its words, a risk of unauthorised command execution. Gateway admins can register stdio MCP servers, which the docs say should be treated as code execution on the host.
The sizing guide is honest about weight: 4 vCPU and 8 GB RAM to evaluate, 8 vCPU and 16 GB for a shared server, before any local model. Against OpenHands, which is narrower and tuned for software tasks, DeerFlow is the broader generalist; against Hermes Agent it is heavier but more of a team service. For a group that wants one self-hosted agent people message from Slack and leave running for an hour, it is the most complete open option we have read about, and the daily commit pace means the configuration you learn this month will shift.
Where it earns its place
A five-person market research team
Turn a question into a sourced report with charts overnight
The lead agent plans the research, sub-agents search and fetch through Tavily or Brave, and code in the Docker sandbox builds charts. The software is free; cost is model tokens plus an 8 vCPU, 16 GB server.
A platform engineer at a 200-person company
Give staff one internal agent they can message from Slack or Feishu
Run the Docker stack on Linux, keep it on loopback behind an authenticating nginx proxy and IP allowlist, and enable the Slack channel in Socket Mode. Budget time for the security setup the README insists on.
A developer already using Claude Code
Hand off long research jobs without leaving the terminal
Install the claude-to-deerflow skill so Claude Code sends tasks to a running DeerFlow instance and checks their status, keeping the hours-long work off the coding session.
Before you commit
- Setup
- Clone the repo and run make setup, a wizard of about two minutes that picks a provider, search and sandbox mode; then make docker-start. Needs Python 3.12+, Node 22+ and Docker Compose 2.24+.
- Integrations
- OpenAI-compatible models with presets for OpenAI, Anthropic, DeepSeek, Qwen, Doubao, Kimi, Zhipu and MiniMax; MCP servers; Telegram, Slack, Discord, Feishu, DingTalk and WeCom; LangSmith and Langfuse tracing; RAGFlow for private knowledge.
- Data and privacy
- Self-hosted; prompts and files go only to the model and search providers you configure. Binds to 127.0.0.1 by default. MIT licence; no GitHub security advisories published, but the README's own security notice warns against untrusted networks.
- Cost at scale
- No paid edition is described. The README sizes a shared server at 8 to 16 vCPU and 16 to 32 GB RAM; model tokens for multi-agent, hours-long runs are the larger bill.
- Limits
- About 990 open issues; 2.0 broke compatibility with v1 entirely; configuration changes often; macOS and Windows are suited to evaluation, Linux plus Docker to servers.
What’s good
- Sub-agents, sandbox, memory, skills and scheduling ship together
- Docker, local or Kubernetes sandboxes; loopback-only by default
- Model-agnostic with presets for eight major providers
- IM channels, Gateway API, Python client and TUI
What’s not
- Executes commands; unsafe to expose without an auth proxy and allowlist
- Sprawling configuration; the README exceeds 4,000 lines
- Needs 8 vCPU and 16 GB RAM for a shared server
- About 990 open issues and a fast-moving codebase
Best for
Teams that want a self-hosted, long-running agent service for research, reports and code jobs and can run it behind proper network controls.
DeerFlow alternatives
DeerFlow: common questions
- Is DeerFlow worth it?
- A serious, batteries-included agent server from ByteDance that does far more than research now, provided you treat it as the code-executing service it is. We rate it 4.2 out of 5.
- How much does DeerFlow cost?
- Free and open source (MIT); you pay for the model API and the server it runs on.
- What are the downsides of DeerFlow?
- Executes commands; unsafe to expose without an auth proxy and allowlist; sprawling configuration; the README exceeds 4,000 lines; needs 8 vCPU and 16 GB RAM for a shared server; about 990 open issues and a fast-moving codebase.
- Who is DeerFlow best for?
- Teams that want a self-hosted, long-running agent service for research, reports and code jobs and can run it behind proper network controls.
- How much does DeerFlow cost for a team or in production?
- No paid edition is described. The README sizes a shared server at 8 to 16 vCPU and 16 to 32 GB RAM; model tokens for multi-agent, hours-long runs are the larger bill.
- Is DeerFlow safe for company data?
- Self-hosted; prompts and files go only to the model and search providers you configure. Binds to 127.0.0.1 by default. MIT licence; no GitHub security advisories published, but the README's own security notice warns against untrusted networks.
- What are the best DeerFlow alternatives?
- The closest alternatives we have reviewed are OpenHands, Hermes Agent, LangGraph, Magentic-UI (MagenticLite) and Agent Zero.
Answers checked Oct 31, 2026.
Make DeerFlow? Show this rating on your site
