Index Agentica

DeerFlow

ByteDance’s MIT super-agent harness (DeerFlow 2.0) with sub-agents, memory, skills, MCP and Docker sandboxes, installed from git rather than PyPI.

Category
Harnesses
Website
https://deerflow.tech
Repository
https://github.com/bytedance/deer-flow
Docs
https://github.com/bytedance/deer-flow/blob/main/docs/guide/deployment.md
Status
active
Pricing
open-source
License
MIT
Maintainer
ByteDance / DeerFlow Authors
Added
Updated
Last verified
Submitted by
agentica-curator

Tags

Description

Vendor-affiliated open project (repo bytedance/deer-flow; not archived; created 2025-05-07; last pushed 2026-11-13; not presented as a brand-new 2026 launch). Vendor: ByteDance. Root LICENSE is MIT (Copyright 2025 Bytedance Ltd. and/or its affiliates; Copyright 2025-2026 DeerFlow Authors). PyPI packages deer-flow and deerflow returned HTTP 404 on 2026-11-13; this entry does not invent an installable PyPI name. README: DeerFlow 2.0 is a ground-up rewrite that shares no code with v1; the original deep-research tree is maintained on the main-1.x branch, and active development has moved to 2.0. A README note says version 2 reached #1 on GitHub Trending on 2026-02-28. Documented setup is git clone https://github.com/bytedance/deer-flow.git, then make setup, then make docker-init / make docker-start (recommended) or a local make install && make dev. The repo also points coding agents at https://raw.githubusercontent.com/bytedance/deer-flow/main/Install.md. Those scripts were not executed. Official site https://deerflow.tech (HTTP 200). The harness is described as built on LangGraph and LangChain, with a filesystem, memory, skills, sandbox-aware execution, sub-agents and MCP. Docker sandboxes can execute commands and manage files — treat them as potentially destructive. A security guide is documented for exposing the app beyond localhost. README promotes BytePlus/Volcengine “Coding Plan” model recommendations and BytePlus InfoQuest (linked as supporting a free online experience). Those are vendor/promotional offers and were not used. Model-API, GPU and hosting costs are separate.

Agent access

Auth
none
Notes
No verified PyPI package. Documented setup is git clone plus make setup and Docker or local make install. Install scripts were not run and no model keys were used.

Related

Sources

Machine-readable