An open-source AI agent platform published by ByteDance. It aims to carry large jobs through to completion — writing research reports, implementing code, producing documents and content — that take anywhere from minutes to hours, without a person having to steer every step. It first appeared in May 2025 as a framework specialized in “Deep Research,” but version 2.0, released in 2026, is a ground-up rewrite that expands it into a general-purpose agent harness with sub-agents, memory, skills, and a sandbox. It is MIT-licensed and can be run on your own machine or server.
Key Features
- Autonomous long-running tasks: It breaks a given goal into a plan and runs the sub-tasks in parallel or in sequence. It evaluates its own progress along the way and reorganizes the remaining work when needed, so it can take on jobs too large for a single chat exchange
- Division of labor through sub-agents: You can spin up child agents with separate roles — a researcher, a coder — and assign each a portion of the work. The parent agent collects and integrates the results
- Long- and short-term memory with context compaction: Context is retained across conversations and tasks, and long histories are automatically summarized and compressed, so information is less likely to be lost during extended work
- Extensible through skills and tools: A skill takes the form of a Markdown file describing a workflow and its best practices. Research, report writing, and slide creation come built in, and you can add your own
- MCP support for external tools: Register Model Context Protocol (MCP) servers in a config file and their capabilities appear alongside the built-in tools. When too many tools are exposed, it can load only the ones needed for the task
- Docker-based sandbox: Code execution, shell commands, and file operations run in an isolated environment. An all-in-one container bundles a browser, shell, file system, and VSCode Server
- A choice of models: It supports Doubao, DeepSeek, OpenAI, Gemini, Qwen, and Claude, among others, plus access via OpenRouter and local models through vLLM, so you can match the model to the task and the budget
- Use it from chat tools: Channel integrations cover Telegram, Slack, Feishu, WeChat, and DingTalk, letting you hand off tasks from the chat app you already use
Pricing
| Plan | Cost | What it covers |
|---|---|---|
| Self-hosted (open source) | $0 | All features. MIT-licensed, commercial use permitted |
| LLM usage | Metered by each model provider | Separate API costs for OpenAI, Gemini, DeepSeek, Doubao, and others |
| Runtime environment | Server costs | Roughly 4 vCPU / 8 GB RAM for local evaluation, around 16 vCPU / 32 GB RAM for a shared server |
Pricing is current as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- The software itself is free, and the MIT license keeps commercial use and modification straightforward
- Because it runs in your own environment, your data does not have to be handed to an external SaaS
- You are not locked to one model and can switch between performance-first and cost-first options
- Sub-agents, memory, sandbox, and skills are all included, so you get running faster than building an agent platform yourself
- MCP support lets you connect the tooling you already have
⚠️ Cons
- Setup assumes familiarity with Node.js, Python, and Docker, making it hard for a non-engineer to deploy alone
- Server and LLM API costs are on you, and token consumption grows the longer your tasks run
- Version 2.0 is a full rewrite with no compatibility with v1, so v1-era guides and configurations no longer apply
- The quality of autonomous execution depends heavily on the model you connect, and verifying the output remains the user’s responsibility
- Adding tools that overlap with built-in ones, such as MCP filesystem servers, can make behavior unstable
Comparison with Similar Services
| Criteria | DeerFlow | OpenHands | Manus | Dify |
|---|---|---|---|---|
| Provider | ByteDance | All Hands AI | Butterfly Effect (Monica) | LangGenius |
| Type | General-purpose agent harness | Agent specialized in software work | Cloud-hosted autonomous agent | LLM app development platform |
| Deployment | Self-hosted | Self-hosted / cloud | Cloud only | Self-hosted / cloud |
| Source code | Open (MIT) | Open | Closed | Open |
| Strengths | Long-running research, coding, and content tasks | Writing and fixing code | Hands-off general tasks | Building workflows and chatbots |
| Cost model | Free core + API and server costs | Free core + API and server costs | Paid subscription | Free core or paid cloud + API costs |
Who Is It For
- People who want to hand off multi-step work, from research through to a finished document
- Companies that cannot send internal or customer data to an external service and need agents running in-house
- Developers who want to compare models in operation rather than commit to a single AI vendor
- Teams that would rather add their own skills on top of a proven foundation than build an agent platform from scratch
- Anyone who has already set up MCP servers and is looking for a runtime that puts them to use
Summary
DeerFlow is an open-source project from ByteDance, rebuilt from a Deep Research framework into a general-purpose agent harness. It ships with the pieces long-running tasks require — sub-agents, memory, skills, and a sandbox — and gives you a wide choice of models. In exchange, getting started calls for developer-level knowledge and a server environment; this is not a cloud service you can sign up for and use immediately. If you want agents running in your own environment and want to pick the model yourself, it is a strong candidate.