DeerFlow

DeerFlow (Deep Exploration and Efficient Research Flow) is a super agent harness that orchestrates sub-agents, memory, and sandboxes to carry out tasks that run from minutes to hours: research reports, data analysis, slide decks, web pages, and image or video generation. Version 2.0 is a ground-up rewrite of the original Deep Research framework. Built on LangGraph and LangChain, it ships with a filesystem, memory, skills, sandbox-aware execution, and the ability to plan and spawn sub-agents for multi-step tasks.

Capabilities come from skills: Markdown files that define a workflow, best practices, and references to supporting resources. Built-in skills cover research, report generation, slide creation, web pages, and image and video generation, and you can add your own, replace the built-in ones, or combine them into compound workflows. Skills load progressively, only when a task needs them, which keeps the context window small. The Capability Center manages skills alongside plugins (MCP servers and Lark/Feishu integration), and skill archives export from one DeerFlow instance and import into another.

Code runs in a sandbox: directly on the host, in Docker containers (the All-in-One Sandbox combines browser, shell, file, MCP, and VS Code Server in one container), in Kubernetes pods through a provisioner service, or on E2B. Each sub-agent runs in its own isolated context, and within a session DeerFlow summarizes completed sub-tasks, offloads intermediate results to the filesystem, and compresses what is no longer relevant. An optional agentic browser tool group, powered by Playwright, keeps a live per-conversation browser session so the agent can navigate, click, type, and submit forms on JavaScript-heavy sites.

DeerFlow is model-agnostic and works with any LLM that implements the OpenAI-compatible API. It carries adapters for DeepSeek reasoning models, Anthropic extended thinking, Codex CLI and Claude Code OAuth credentials, OpenRouter, vLLM, and Ollama, and it recommends models with long context windows, reasoning, multimodal input, and tool use for function calling and structured outputs. External coding agents such as MiniMax Code attach over ACP (Agent Client Protocol), and a claude-to-deerflow skill lets Claude Code send tasks to a running instance.

Beyond the web workspace, DeerFlow receives tasks from Telegram, Slack, Discord, Feishu/Lark, DingTalk, WeChat, WeCom, and QQ channels, exposes a LangGraph-compatible Gateway API and an embedded Python client, and ships a terminal workbench that runs embedded without the Gateway or frontend. Long-term memory uses the DeerMem backend by default with OpenViking as an optional alternative, private knowledge retrieval cites RAGFlow evidence, and LangSmith, Langfuse, and Monocle tracing are built in. Projects group conversations under shared instructions and a document shelf, and scheduled tasks run prompts once or on a recurring schedule.