DeerFlow 2.0: ByteDance releases open-source alternative to Manus

What it's about

ByteDance, TikTok's parent company, has released DeerFlow 2.0, an open-source framework for AI agents. The name stands for "Deep Exploration and Efficient Research Flow." An agent here is an AI system that doesn't just output text but carries out tasks on its own: it researches, writes and runs code, and creates files. ByteDance calls DeerFlow a "Super Agent Harness," a framework that holds multiple agents, tools and memory together. Reports present it as a free alternative to the commercial service Manus. The code uses the MIT license, which largely allows free use, modification and redistribution.

How DeerFlow works

According to MarkTechPost, DeerFlow doesn't just live in a chat window. The agent runs in a real, isolated Docker container (a sealed-off virtual work environment) with a file system and a Bash terminal. Given a task like "analyze this CSV file," it doesn't just suggest a script. It sets up the environment, installs dependencies, runs the code and delivers, say, a finished chart.

For complex jobs, such as "Research the ten most important AI startups of 2026 and build a presentation," a lead agent acts like a project manager:

• It breaks the task into subtasks.
• Sub-agents run in parallel, for example for web research, competitor analysis or images.
• The results come together at the lead agent.
• A final step produces the end product, such as slides or a web app.

Heise adds that each sub-agent has its own context, its own tools and its own completion conditions. That is what makes parallelization possible. To keep the context small and save tokens, the system summarizes continuously. Under the hood, DeerFlow is built on LangChain and LangGraph, two widely used libraries for agent workflows.

Skills, memory and folder structure

Users define workflows through "skills," which are Markdown files with instructions. Even the research workflow is a skill. According to heise, DeerFlow loads skills only when needed, which saves memory and compute time. "Tools" are more complex functions such as web search or the secure execution of Bash and Python code.

DeerFlow also has long-term memory. It remembers preferences, writing style and project context across sessions and, according to heise, stores this data locally. Geeky Gadgets describes three folders: Uploads for inputs, Workspace as a temporary work area and Outputs for results. As an example, the report cites an analysis of Titanic passenger data in which DeerFlow wrote its own Python scripts, created visualizations and fixed errors.

Origins

According to MarkTechPost, DeerFlow was originally a specialized tool for deep research. But internal users used it to build data pipelines, dashboards and entire web apps. ByteDance then rewrote the framework from scratch. Heise reports that the developers recognized LangChain's extensibility as the decisive factor.

What you need to try it

Geeky Gadgets names four CPU cores and 8 GB of RAM as the minimum. For local language models, you will usually also need a graphics card. With cloud models, your data goes to external providers. According to heise, installation runs through a YAML configuration file and Docker containers. They open port 2026. All environment variables belong in a .env file, because the system doesn't use shell variables. The instructions are on GitHub. Linux, Mac and Windows with Docker should work, but the sources don't state this explicitly. DeerFlow itself is not a model and needs little storage space, while a local language model needs considerably more.

Models, search and security

DeerFlow is model-agnostic. It uses any OpenAI-compatible API, so according to MarkTechPost that includes GPT-4, Claude, Gemini or local models via DeepSeek and Ollama. Heise writes that ByteDance wants to steer users toward its own coding subscription, but that isn't required. Web search is harder when you run locally. You can plug in Tavily (1,000 requests per month for free) or ByteDance's proprietary service InfoQuest (1,000 requests in total per account, pricing for more not public). According to heise, ByteDance doesn't offer this service in China at all.

Geeky Gadgets points out that the default local execution runs in a sandbox without secure isolation, which can be risky in multi-user environments. Docker or Kubernetes improve this but require expertise. Heise explicitly sees the sandboxes as a response to dangers like those seen with OpenClaw, which according to the report has in some cases stolen credit cards or taken over computers.

Our take

DeerFlow is technically interesting: parallel sub-agents, a sandbox per process and skills as Markdown are practical ideas. Heise sees tools like n8n or Dify as considerably more static. Compared with Manus, a managed service with credit-based pricing, DeerFlow offers more control but demands more work from you. According to Geeky Gadgets, there is no comparison of cost or performance. Weaknesses: slides are sometimes produced as static images and are hard to edit, and web scraping depends on what sites allow. The closeness to ByteDance products such as DeepSeek V3.2 is also read as an ecosystem strategy. The sources contain no independent benchmarks. A report that ByteDance has joined the patent community Open Invention Network has nothing to do with DeerFlow and is left out here.