Overview
DeerFlow is an open-source super agent harness built by ByteDance that orchestrates sub-agents, long-term memory, and sandboxed execution environments to handle complex, multi-step tasks that can span minutes to hours. Version 2.0, released in early 2026, transformed it from a focused deep research framework into a general-purpose AI agent platform capable of conducting research, writing and executing code, building websites, generating presentations, and coordinating multiple sub-agents in parallel.
Powered by LangGraph and LangChain, DeerFlow works with any OpenAI-compatible model — from DeepSeek and Kimi to Claude and GPT — and deploys in minutes via Docker. The entire codebase is MIT-licensed and freely self-hosted, making it a compelling alternative to proprietary agent platforms for teams that need full control over their infrastructure.
DeerFlow reached #1 on GitHub Trending on February 28, 2026 following the version 2 launch.
Key Features
Sub-agent orchestration — The lead agent dynamically spawns specialized sub-agents that run in parallel, each with its own scoped context, tools, and termination conditions. Results are synthesized back into coherent outputs — enabling research, code, and content tasks that would overwhelm a single-agent context window.
Sandboxed code execution — DeerFlow supports three execution modes: local, Docker, and Kubernetes via a provisioner service. In Docker/Kubernetes modes, tasks run in isolated environments with dedicated filesystems.
Long-term memory — DeerFlow maintains persistent local memory, stored in
memory.jsonby default, to retain preferences and context across sessions. Memory is locally controlled and avoids duplicate accumulation through structured deduplication.Extensible skills system — Built-in skills cover web research, report generation, slide creation, image and video generation, and Python/Bash execution. Skills load progressively — only when the task needs them — keeping context windows lean. Custom tools can be added via MCP servers or Python functions.
Multi-model support — Configurable via
config.yaml, DeerFlow works with any OpenAI-compatible provider including DeepSeek, Kimi, Claude, Gemini, and self-hosted models. Supports LangChain providers and CLI-backed models such as Claude Code OAuth.IM channel integrations — Connect DeerFlow to Telegram, Slack, or Feishu/Lark with long-polling or WebSocket transports, enabling agents to receive tasks and deliver results through existing team communication channels.
How to Get Started
Prerequisites: Docker and Docker Compose (recommended), or Python 3.12+ and Node.js 22+ for local development.
Docker Quick Start (recommended):
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make config # generates config.yaml
# Edit config.yaml: add model provider and API keys
make docker-init
make docker-start
Access the interface at http://localhost:2026.
Local Development:
make check # verify system requirements
make install # install Python and Node.js dependencies
make dev # start frontend and backend concurrently
Model Configuration:
Open config.yaml and define your model provider (OpenAI, Anthropic, OpenRouter, DeepSeek, etc.) with the API key. DeerFlow recommends models with long context windows (100K+ tokens), reasoning capabilities, and strong tool-use. Set secrets in a .env file rather than committing them to the config.
Community & Ecosystem
DeerFlow has attracted significant community interest since the version 2 launch. The GitHub repository surpassed 25,000 stars within weeks of the v2.0 launch, reflecting strong interest from AI researchers, application developers, and enterprise teams evaluating self-hosted agent infrastructure.
The project is actively maintained by a core team including contributors Daniel Walnut and Henry Li. Contributions are welcomed through standard GitHub pull requests, and the repository includes comprehensive documentation covering architecture, configuration, and development guides.
The MCP server support means DeerFlow can integrate with a rapidly expanding ecosystem of MCP-compatible tools and services, extending its capabilities without modifying the core codebase. Telegram, Slack, and Feishu integrations make it easy to embed DeerFlow into existing team workflows.
Pricing & Plans
DeerFlow is completely free and open-source under the MIT license. There are no usage fees, subscription tiers, or credit limits — infrastructure costs (cloud hosting, model API usage) are the only expenses.
| Component | Cost |
|---|---|
| DeerFlow software | Free (MIT license) |
| Self-hosted deployment | Variable — your own infrastructure costs |
| Model API usage | Varies by provider (e.g., DeepSeek, Anthropic, OpenAI) |
Best For
- AI researchers and engineers who want full control over agent infrastructure and want to study or extend the orchestration logic
- Developer teams building internal automation tools that need sandboxed code execution without exposing data to third-party SaaS platforms
- Organizations evaluating self-hosted alternatives to proprietary agent products like Pi Monorepo or commercial agent APIs
- Startups and builders who need a production-grade agentic framework without per-seat or per-call licensing costs
- Teams already using Slack, Telegram, or Feishu who want to embed an AI agent natively in their communication workflow
FAQ
What is DeerFlow 2.0?
DeerFlow 2.0 is a complete rewrite of ByteDance's original DeerFlow deep research framework. Version 2.0 expands the scope from research automation to a general-purpose super agent harness that can orchestrate sub-agents, execute sandboxed code, maintain persistent memory, and integrate with messaging platforms.
Is DeerFlow free to use?
Yes. DeerFlow is MIT-licensed and completely free. You host it yourself on your own infrastructure. The only costs are cloud compute (if you deploy to a VPS or cloud) and model API usage from your chosen provider.
What AI models does DeerFlow support?
DeerFlow works with any OpenAI-compatible API. This includes OpenAI, Anthropic (Claude), DeepSeek, Kimi, Gemini via OpenAI-compatible endpoints, and self-hosted models through OpenRouter or similar proxies. Model configuration is done in config.yaml.
How does sub-agent orchestration work?
The lead agent analyzes a task, breaks it into subtasks, and spawns specialized sub-agents for each. Sub-agents run with their own scoped context and tool access, potentially in parallel. When finished, they return structured results to the lead agent, which synthesizes a final output. This architecture enables tasks too large or complex for a single context window.
Is it safe to run DeerFlow with code execution enabled?
DeerFlow includes sandboxed execution via Docker and Kubernetes to isolate agent activities from the host system. Security experts recommend deploying with hardened container images, restricted network access, and principle-of-least-privilege configurations. Code-executing agents carry inherent risks and should not be run with unrestricted system access in production.
Does DeerFlow require internet access?
Web research and external API calls require internet access. The core agent runtime and sandboxed execution can operate offline if configured with locally hosted models and no web search skills.
How does DeerFlow compare to other autonomous agents like Manus?
DeerFlow is self-hosted and MIT-licensed, giving teams direct control over data, infrastructure, and extensions. For a fair comparison with Manus or other autonomous-agent products, verify deployment model, pricing, and customization options from each product's official documentation. DeerFlow is generally better suited for technical teams who want to customize and extend the agent stack.
Where can I find documentation and support?
Full documentation is available in the GitHub repository at github.com/bytedance/deer-flow. The project includes architecture guides, configuration references, and contribution guidelines. Community support is available via GitHub Issues and Discussions.
