deer-flow

DeerFlow (Deep Exploration and Efficient Research Flow) is an open-source "super agent harness" by ByteDance. It coordinates sub-agents, long-term memory, sandboxed execution, and extensible skills to handle long-horizon tasks that may take minutes to hours, such as deep research, coding, and content creation. Version 2.0 is a ground-up rewrite and does not share code with the 1.x branch; the original deep-research framework remains maintained separately. The project provides a setup wizard (make setup) that generates a minimal config.yaml and .env, a make doctor diagnostic, and a support-bundle command for issue reporting. It can be run with Docker Compose for development or production (make docker-start, make up) or locally (make dev) with Node.js 22+, pnpm, uv, and nginx. Deployment sizing guidance covers local evaluation, Docker development, and long-running servers. It supports many LLM providers via LangChain OpenAI-compatible ChatOpenAI, OpenAI Responses API, CLI-backed providers (Codex CLI, Claude Code OAuth), vLLM for local models, and OpenRouter gateways. Per-model settings include reasoning/thinking toggles, effort levels, API keys, base URLs, and pricing. Core features include an extensible Skills & Tools system with Claude Code integration, session goals, manual context compaction, sub-agents, sandbox and file-system isolation, context engineering, and long-term memory. It also offers MCP server support, IM channel integration, optional tracing with LangSmith, Langfuse, or Monocle, use of multiple providers, personal access tokens, scheduled tasks, an embedded Python client, and a terminal workbench (TUI). A web UI is served on http://localhost:2026 with browser login using HttpOnly session cookies; production multi-worker deployments require Postgres and a Redis stream bridge. The README includes a security notice warning that improper deployment may introduce security risks and provides recommendations. It recommends running with models such as Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5, and integrates BytePlus InfoQuest for web/image search and crawling. The project is MIT-licensed and includes contribution guidelines and documentation links.