DeerFlow

Free

DeerFlow is an open-source AI agent framework. It builds systems that run for hours, not seconds.

You write a goal in plain words. It returns finished files: reports, code, charts, videos.

It sits in AI agents work. You self-host it with Docker.

Open it if you need agents that run code and return files. Skip it if you want a chat bot or a hosted service.

You need this if your work spans hours and needs real execution.

• Teams building multi-hour agent workflows: Tasks run from minutes to hours. Sub-agents work in parallel without you watching.
• Developers comfortable with self-hosted setup: You run Docker, YAML, and CLI. The control is worth the work.
• Organizations needing data sovereignty: Everything stays local. No seat fees. No cloud lock-in.
• Teams wanting end-to-end execution: Agents write code, run it, and hand you files. Not suggestions. Output.
• Users needing sandboxed security: Each agent runs in its own Docker container. Files, shell, and browser are isolated.
• Developers building custom skills: You write workflows in Markdown. Reuse them across projects.
• Teams evaluating agentic systems for research: It handles literature reviews, data pulls, and report writing in one pass.
• Businesses wanting production-ready execution: Kubernetes and GPU support exist. You can ship without building from scratch.

These are the parts you notice after you get it running.

• Real Docker sandbox: Agents do not pretend to run code. They actually execute it. You get files you can download and use.
• Lead agent architecture: One agent breaks your goal into pieces, then farms them out. You do not write the orchestration.
• Skills in plain Markdown: You define workflows without learning new formats. It feels like writing docs, not config.
• Model-agnostic design: Plug in OpenAI, local models, or others. You control the routing and the budget.
• Persistent memory: Agents remember state across sessions. Long jobs do not start from zero.

You need Docker and some patience.

1. Clone the repo and run `make setup` to pick your LLM provider.
2. Configure your agents through YAML files.
3. Write or pick a skill that matches your task.
4. Submit your goal through the web UI or CLI.
5. Wait. Download your files when the run finishes.

Skip it if this sounds like too much.

• Complex setup: Expect 30 minutes to get running. Hosted tools open faster.
• Security burden: You audit the sandbox. You manage secrets. The risk is yours.
• Heavy resource use: Multi-agent local workloads need real compute.
• Fewer skills out of box: You may need to build more than you want.
• ByteDance provenance: Some orgs will review this. Factor it in if you are regulated.