DeerFlow
Sub-agents, memory, skills, and sandboxes carry research and coding tasks across hours.
What it is
DeerFlow is an open-source SuperAgent harness built to carry out research, coding, and content-production tasks that may run for minutes to hours. Once you give it a goal, it can split the work across multiple sub-agents, run commands, read and write files, and call tools inside isolated sandboxes, while context compression and long-term memory keep the task going across stages, and finally deliver a report, code, or other output.
A typical run starts by configuring the model, search services, and execution permissions, then submitting a task from the web or the terminal and waiting for each stage to finish. Unlike the usual single-turn question-and-answer tool, its focus is not producing an immediate reply but organizing multi-step execution; 2.0 is a complete rewrite of the earlier Deep Research framework, and its positioning has shifted toward full task orchestration.
How to run
The repository provides an interactive configuration wizard. The shortest local startup flow is:
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make setup
make devmake setup generates config.yaml and writes secrets into .env. For a formal deployment you can also use Docker: run make up, then visit http://localhost:2026 to enter the web interface.
Core mechanisms
- Sub-agent orchestration: the main task can be split across multiple sub-agents, with an upper limit on how many run at once to control resource consumption.
- Isolated execution: commands, files, and tool operations are carried out in a sandbox, and permissions can be tightened during configuration.
- Context continuity: session goals, context compression, and long-term memory together preserve the state of long-running tasks.
- Skill extensions: skills and tools define what the agent can do, and it can also connect to MCP, messaging channels, and different model providers.
Who it is for
DeerFlow suits teams that want to deploy their own research and coding agents, especially in scenarios that require generating reports, working on repositories, or completing multi-stage tasks in sequence. The web interface, terminal workspace, and embedded Python client cover manual operation and programmatic integration respectively.
Usage boundaries
The project uses the MIT license, but a usable large model still has to be configured before running, and external model and search services may charge fees. For development, at least 4 vCPU, 8 GB of memory, and 20 GB of free disk space are recommended; a long-running server is better suited to Linux with Docker. Local development on Windows only supports Git Bash, not native cmd.exe or PowerShell.
Opening up Bash or file-write permissions increases sandbox risk, so permissions should be restricted when handling untrusted tasks. Running multiple Gateway workers in production also depends on Postgres, Redis, and extra ownership configuration.