DeerFlow

// Overview

DeerFlow — short for Deep Exploration and Efficient Research Flow — is an open-source agent harness from ByteDance, released under the MIT license. It started life as a deep-research framework, but the community pushed it into data pipelines, slide decks, dashboards and content workflows, so the maintainers rewrote it from scratch as DeerFlow 2.0: a runtime that gives an agent the infrastructure it needs to actually finish work, rather than a framework you wire together yourself.

Built on LangGraph and LangChain, DeerFlow 2.0 ships with a filesystem, long-term memory, extensible skills, sandbox-aware execution and the ability to plan and spawn sub-agents for multi-step tasks. Skills are the extension point: each one is a Markdown module describing a workflow and its supporting resources, loaded progressively only when a task needs it so the context window stays lean. Built-in skills cover research, report generation, slide creation, web pages, and image and video generation, and you can add, replace or combine your own.

The project is model-agnostic — anything speaking an OpenAI-compatible API works, including CLI-backed providers such as Codex CLI and Claude Code — and runs as a web app, a terminal workbench (TUI), an embedded Python client, or an MCP server. Note that DeerFlow 2.0 shares no code with version 1; the original deep-research framework lives on the `main-1.x` branch.

// What it does

• Extensible Markdown skills, loaded progressively and activatable per turn with a `/skill-name` prefix, with `allowed-tools` policies that scope what an active skill may call.
• Sub-agents the lead agent can spawn on demand, each with its own scoped context, tools and termination conditions, reporting structured results back for synthesis.
• Sandbox-aware execution with local, Docker/AIO and E2B sandbox providers, plus a shared filesystem for offloading intermediate results.
• Long-term memory across sessions via the default local DeerMem backend, with optional mem0 and OpenViking backends.
• Context engineering built in: isolated sub-agent contexts, aggressive summarization of completed sub-tasks, and manual `/compact` compaction from the composer.
• Session goals (`/goal <condition>`) that persist across turns until DeerFlow judges them satisfied, plus scheduled tasks and IM channel integrations.
• Tracing hooks for LangSmith, Langfuse and Monocle, and an MCP server so other clients can drive the harness.

// Getting started

DeerFlow needs Python 3.12+ and Node.js 22+, and is easiest to run with Docker. Clone the repo, run the setup wizard to pick an LLM provider and safety preferences, then start the services. `make doctor` verifies your setup at any point.

git clone https://github.com/bytedance/deer-flow.git
cd deer-flow

make setup
make doctor   # verify the setup and get fix hints

make docker-init    # pull the sandbox image (once)
make docker-start   # start services
make docker-logs    # follow logs

make check     # verifies Node.js 22+, pnpm, uv, nginx
make install   # backend + frontend deps + pre-commit hooks
make dev

/goal finish the implementation and make all tests pass
/goal          # show the active goal
/goal clear    # clear it

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Run long-horizon research that spans minutes to hours, with sub-agents fanning out over independent read-only sources and the lead agent synthesizing the result.
• Turn a prompt into a deliverable — a report, a slide deck, a web page, or generated images and video — using the built-in skills rather than hand-built prompt chains.
• Give a self-hosted agent a persistent memory of your stack, preferences and recurring workflows, stored locally and under your control.
• Package your own team workflows as Markdown skills so an agent can pick them up progressively instead of carrying every instruction in context.

// How DeerFlow compares

DeerFlow alongside other open-source agent frameworks & builders tools AI/TLDR tracks, ranked by GitHub stars.