RepoDaily · 2026-06-23 · AI Agents

DeerFlow 2.0: ByteDance's Open-Source SuperAgent Harness for Research, Code, and Creation

#7 Self-hosted app Python +736 bytedance/deer-flow Open repository

A ground-up rewrite that orchestrates sub-agents, sandboxes, memory, and extensible skills into a single self-hosted agent runtime — with a LangGraph-compatible gateway and Docker-first deployment.

Repo typeSelf-hosted app
Best forDevelopers and teams who want a self-hostable multi-agent harness for deep research, coding, and content creation without relying on a hosted SaaS agent platform.
Risk levelMedium
Time to evaluate1–2 hours with Docker; longer for sandbox and multi-provider configuration.

Primary question: Does your team need a self-hosted, extensible agent harness that can coordinate sub-agents, execute code in sandboxes, and retain long-term memory?

83/100

RepoDaily adoption score

RepoDaily rates this as 83/100 (strong) for adoption: evidence, installation path, production risk, differentiation, license clarity, and AI/agent fit are scored from the article sources and adoption notes.

Directional score from RepoDaily sources and adoption notes, not a benchmark.Risk: Medium
76Evidence quality

4 source(s) across 2 source category/categories, plus a RepoDaily-specific evidence module when available.

100Installability

5 workflow step(s), 4 next-action step(s), and 5 command/install signal(s) were detected.

66Maintenance confidence

Trending momentum is +736 stars, with maintenance/release/issue signals counted when present.

88Production readiness

Risk is marked medium, with 4 security note(s) and 3 explicit skip condition(s).

80Differentiation

2 opportunity lens item(s), 3 alternative(s), and 0 type-specific section(s) support differentiation.

82License clarity

License source or license wording is present.

90Agent / AI fit

9 AI/agent-related signal(s) were detected in the article text and metadata.

Project overview

DeerFlow (Deep Exploration and Efficient Research Flow) is ByteDance's open-source super-agent harness. Version 2.0 is a complete rewrite that shares no code with the original v1 deep-research framework — the old branch remains available, but active development has moved entirely to 2.0.

The system orchestrates sub-agents, persistent memory, code-execution sandboxes, and a pluggable skills system to handle long-horizon tasks spanning research, coding, and content creation. It exposes a LangGraph-compatible API gateway, a web frontend, and an embedded Python client, making it deployable as both an interactive app and a programmatic backend.

DeerFlow runs as a self-hosted application behind nginx, with a Docker-first deployment story. It supports multiple LLM providers — including OpenAI-compatible endpoints, OpenRouter, vLLM, Codex CLI, and Claude Code OAuth — and integrates BytePlus's InfoQuest for intelligent search and crawling.

Problem it solves

  • Most agent frameworks are either thin orchestration layers (no sandbox, no memory) or hosted SaaS products you can't self-host or audit.
  • Long-horizon tasks — multi-step research, iterative coding, content pipelines — require coordinated sub-agents, persistent context, and safe code execution that few open-source tools provide together.
  • Configuring multiple LLM providers, search backends, and execution safety policies is usually a manual, error-prone process.

How it works

  1. Clone the repo and run `make setup` — an interactive wizard guides provider selection, web search, sandbox mode, bash access, and file-write preferences, generating config.yaml and .env.
  2. Run `make docker-init` then `make docker-start` to launch nginx, frontend, gateway, and optional provisioner containers with hot-reload enabled.
  3. Access the web interface at localhost:2026 or interact programmatically via the Gateway API and LangGraph-compatible endpoints.
  4. The agent runtime orchestrates sub-agents, consults long-term memory, executes tasks in sandboxed environments, and invokes extensible skills to complete multi-step objectives.
  5. Use `make doctor` at any time to verify configuration and receive actionable fix hints.

Deployment & Architecture

DeerFlow runs behind nginx on port 2026, which routes non-API requests to the frontend (port 3000) and API requests to the Gateway (port 8001). The Gateway hosts a LangGraph-compatible runtime that handles agent interactions.

Docker is the recommended path: `make docker-init` builds images and installs dependencies; `make docker-start` launches all services with hot-reload. A local development path is also available for those who prefer running services directly with Node.js 22+, pnpm, uv, and nginx.

The README provides concrete host-resource guidance: 4 vCPU / 8 GB RAM as a starting point, scaling to 16 vCPU / 32 GB RAM for shared multi-agent test servers. It explicitly warns that 2 vCPU / 4 GB environments often fail under normal workloads.

Model Provider Flexibility

  • OpenAI-compatible providers via langchain_openai:ChatOpenAI, including OpenRouter with custom base_url.
  • OpenAI Responses API support with use_responses_api flag.
  • vLLM support (0.19.0+) via a custom VllmChatModel provider, including reasoning-model toggles.
  • CLI-backed providers: Codex CLI (reads ~/.codex/auth.json) and Claude Code OAuth (multiple credential paths).
  • README recommends Doubao-Seed-2.0-Code, DeepSeek v3.2, and Kimi 2.5 for running DeerFlow.

Skills, Sub-Agents & Memory

DeerFlow's skills system lives in a top-level `skills/` directory with `public/` and `custom/` subdirectories, letting teams extend agent capabilities without forking the core. An extensions_config template covers MCP and Skills configuration.

Sub-agents can be coordinated through the agent runtime, and a long-term memory layer persists context across sessions. The sandbox and file-system layer provides isolated execution for code-generating tasks, and context-engineering features help manage token budgets across multi-step workflows.

Who should pay attention?

Good fit if

  • Teams evaluating a self-hostable alternative to hosted agent platforms like Devin or Manus.
  • Developers already in the LangChain/LangGraph ecosystem who want a compatible agent runtime.
  • Researchers and engineers running long-horizon tasks that need sandboxed code execution and persistent memory.
  • Organizations with their own LLM deployments (vLLM, OpenAI-compatible gateways) that need an agentic front-end.

Skip for now if

  • Teams looking for a managed, no-infrastructure SaaS agent product.
  • Projects that only need a simple single-turn LLM wrapper without orchestration.
  • Environments with less than 4 vCPU / 8 GB RAM — the README warns these often fail.

Risks and cautions

Medium

DeerFlow 2.0 is a fresh rewrite with strong backing, but its multi-service architecture, sandbox configuration, and resource requirements create a non-trivial operational surface.

  • Running the full stack (nginx, frontend, gateway, optional provisioner) requires Docker expertise and adequate host resources.
  • Sandbox mode and bash/file-write access introduce security considerations that the README explicitly calls out.
  • As a 2.0 ground-up rewrite, production battle-testing at scale is still accumulating.
  • Multiple provider configuration paths (OpenAI, vLLM, Codex CLI, Claude OAuth) increase setup complexity for teams using non-standard backends.
  • The README includes a dedicated security notice warning that improper deployment may introduce security risks.
  • Sandbox mode, bash access, and file-write tools are configurable through the setup wizard — teams should restrict these in shared environments.
  • Credential handling supports multiple paths (env vars, auth files) — sensitive keys should be managed via `.env` and never committed.
  • nginx configuration provides same-origin API routing; split-origin deployments require explicit CORS allowlist configuration via GATEWAY_CORS_ORIGINS.

Alternatives to compare

ApproachWhen to useTrade-off
LangGraph (LangChain)
You want to build custom agent graphs yourself without a pre-packaged harness.Open-source; self-hosted
OpenHands
You need an open-source autonomous coding agent with similar sandbox execution.Open-source; self-hosted
AutoGen (Microsoft)
You want a multi-agent conversation framework focused on agent-to-agent dialogue.Open-source; self-hosted

What this trend reveals

Enterprise Research Automation

Teams can deploy DeerFlow behind a corporate firewall with their own LLM endpoints, giving analysts a self-hosted deep-research agent that handles search, synthesis, and report generation without data leaving the network.

Pilot with a single research team; configure InfoQuest or an internal search provider; measure task completion quality against manual workflows.

Custom Skill Marketplace

The skills directory structure (public + custom) invites teams to build domain-specific skills — financial analysis, legal document review, code auditing — that extend the agent's capabilities for vertical use cases.

Build one custom skill for your domain, test it through the web UI, and evaluate whether sub-agent coordination improves over a monolithic prompt approach.

Best next action

Run the Docker Quick Start

Clone the repo, run the setup wizard, and launch the Docker stack to explore DeerFlow's web UI and agent capabilities on a real task.

  1. Clone: git clone https://github.com/bytedance/deer-flow.git && cd deer-flow
  2. Run `make setup` and follow the interactive wizard to configure your LLM provider.
  3. Run `make docker-init` then `make docker-start`.
  4. Open http://localhost:2026 and try a multi-step research or coding task.

RepoDaily verdict

DeerFlow 2.0 is one of the most ambitious open-source agent harnesses available — a self-hostable, extensible platform that combines sub-agent orchestration, sandboxed execution, persistent memory, and a LangGraph-compatible API. For teams willing to manage its infrastructure, it offers a credible path to owning their agent stack end-to-end.

Sources