ByteDance, the company behind TikTok, has released DeerFlow 2.0 - an open-source framework for running coordinated AI agents on your own hardware. Unlike cloud-dependent agent services, DeerFlow runs entirely locally with Docker, lets you use any compatible LLM, and keeps your data on your own machines.
The project hit #1 on GitHub Trending when it launched on February 28, 2026. It currently has over 31,000 stars and is licensed under MIT, meaning you can use it for anything including commercial projects.
What DeerFlow Actually Does
DeerFlow is a “SuperAgent harness” - infrastructure for running AI agents that can research, code, create content, and coordinate with each other. According to ByteDance’s GitHub repository, the framework “orchestrates sub-agents, memory, and sandboxes to do almost anything.”
The key capabilities:
Multi-agent orchestration. A lead agent can spawn sub-agents that run in parallel, each with scoped context and tools. The lead agent synthesizes their results into coherent output. This matters for complex tasks where different aspects can be worked on simultaneously.
Sandboxed execution. Every agent runs in an isolated Docker container with its own filesystem, bash terminal, and file read/write access. Agents can execute Python scripts and bash commands safely without risking your host system.
Persistent memory. DeerFlow maintains long-term memory that survives session restarts. It stores user preferences, domain knowledge, and task context - meaning agents can build on previous work rather than starting fresh each time.
Extensible skills. Capabilities are delivered as Markdown files with optional frontmatter for versioning. Built-in skills include research with source citations, report generation, slide creation, web page building, and image/video generation.
Why This Matters for Self-Hosting
The YUV.AI technical analysis captures what makes DeerFlow different from frameworks like LangChain or CrewAI: “they treat the execution environment as an afterthought, leaving us to figure out sandboxing, memory persistence, and multi-agent coordination ourselves.”
DeerFlow prioritizes runtime infrastructure over reasoning abstractions. While LangChain handles prompt engineering and chain-of-thought reasoning, DeerFlow handles execution - the Docker containers, the file systems, the coordination between agents.
The comparison: if LangChain is like a web framework handling routing, DeerFlow is like Docker/Kubernetes managing execution and orchestration at lower layers.
For self-hosters, this means you get production-ready agent infrastructure without building it yourself. You still design agent logic and select LLMs, but the hard parts of safe execution and coordination are handled.
Getting Started
Requirements: Python 3.12+, Node.js 22+, Docker, and pnpm.
Clone and configure:
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make config # Generates local configurationEdit config.yaml with your model configuration. DeerFlow uses LangChain class paths, so you can use any OpenAI-compatible provider:
models:
- name: "my-model"
class: "langchain_openai:ChatOpenAI"
base_url: "https://openrouter.ai/api/v1" # OptionalSet your API keys in .env, then run:
make docker-init
make docker-start # Development with hot-reloadAccess the interface at http://localhost:2026.
For production deployment, use make up instead.
Recommended Models
ByteDance recommends Doubao-Seed-2.0-Code, DeepSeek v3.2, or Kimi 2.5 for best results. But since DeerFlow uses standard LangChain integrations, you can use any compatible model including local ones via Ollama.
The framework supports multiple sandbox modes: local execution on your host, Docker containers for isolation, or Kubernetes via a provisioner service for scaling.
Integration Options
DeerFlow includes built-in support for messaging platforms:
- Telegram via Bot API with long-polling
- Slack via Socket Mode
- Feishu/Lark via WebSocket
No public IP required - channels auto-start when configured. There’s also a claude-to-deerflow skill that enables DeerFlow interaction from Claude Code terminal.
MCP (Model Context Protocol) servers are configurable with OAuth support for client_credentials and refresh_token flows.
Limitations
DeerFlow is infrastructure, not a turnkey solution. You need to understand Docker, configure your own models, and design agent logic for your use cases. The 220 open issues on GitHub suggest active development with some rough edges.
The framework targets extended autonomous operations - tasks that might run for hours. For simple single-turn queries, it’s overkill. The value is in complex, multi-step work where coordination and persistence matter.
What This Means
ByteDance releasing DeerFlow under MIT license follows a pattern of major tech companies open-sourcing AI infrastructure. Meta has Llama, Microsoft has various AI tools, and now ByteDance contributes agent orchestration.
For the self-hosted AI community, DeerFlow fills a gap. Running LLMs locally with Ollama is well-established. But running coordinated agent systems - with proper sandboxing, memory, and multi-agent orchestration - has required building infrastructure from scratch or using cloud services.
DeerFlow provides that infrastructure as a starting point. It’s not plug-and-play, but it’s a substantial foundation for anyone wanting to run AI agents locally without depending on cloud providers.
The Bottom Line
DeerFlow 2.0 is a serious open-source option for self-hosting AI agent systems. If you need multi-agent coordination with safe code execution and persistent memory, and you’re comfortable with Docker and LangChain, it’s worth evaluating.