DeerFlow
github.com/bytedance/deer-flow- Category
- AI Agents
- Rank
- No. 40Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- bytedance
- GitHub
- 82.0k stars
- Latest release
- v2.0.0
- Date
About
An open-source super agent harness that orchestrates sub-agents, memory, and sandboxes to handle long-horizon tasks that can take minutes to hours. Supports research, coding, and creative tasks with extensible skills and tool integrations.
What it does
DeerFlow provides a configurable workspace where an agent can search, write files, run shell commands, create artifacts, and delegate work. Users choose model providers, web search, memory, and execution permissions through a setup wizard. It can run locally or through Docker, with browser, terminal, Python client, scheduled-task, and MCP surfaces described in the documentation.
Why it's ranked here
DeerFlow is compelling because it treats agent operation as a deployable system, not merely a chat loop. Setup diagnostics, configurable safety controls, multiple model routes, persistent memory, tracing options, and redacted support bundles address practical operating needs. The breadth also raises the cost of understanding and securing a deployment.
What's good
The setup wizard collects model, search, sandbox, shell, and file-write choices, then produces minimal configuration. A doctor command reports actionable fixes, while support bundles exclude environment secrets, conversation text, and user file contents. Memory storage uses per-fact revisions, recoverable journals, atomic replacement, scoped locks, and immutable migration backups. OpenViking failures can be configured separately for reads and writes.
Tradeoffs
This is a substantial Python and Node application with several services, not a lightweight library drop-in. The documented local starting point is four virtual CPUs, eight gigabytes of memory, and twenty gigabytes of free storage. Persistent memory migration is one-way for older application code and requires an operator backup. The OpenViking adapter targets one credential-bound user and lacks fact editing, import, export, and durable at-least-once delivery.
How to use it well
Use DeerFlow when a team wants a self-hosted agent workspace with controlled execution, provider choice, artifacts, and persistent context. Start with Docker, keep shell and file permissions narrow, run the diagnostic command, and enable optional memory or tracing only after the base flow works. It does not replace a durable job queue, multi-user credential provisioning, or a database-backed memory service.
Technical notes+
README.md documents Python 3.12+, Node.js 22+, Docker and local workflows, model configuration through LangChain classes, OpenAI-compatible gateways, vLLM, Codex CLI, and Claude Code OAuth. Makefile exposes setup, doctor, support-bundle, dependency installation, development, production, Docker, nginx, and static diagnostic targets. docs/OPENVIKING.md specifies a langchain-openviking==0.1.0 adapter, one stable OpenViking Session per DeerFlow thread, bounded hash-and-counter cursors, fail-open reads, configurable write failure behavior, and no at-least-once guarantee. docs/plans/STORAGE_REWRITE_CHANGES.md describes Markdown fact sharding, separate manifest and fact revisions, in-process plus cross-process locks, journaling, bounded conflict retries, atomic replacement, migration backups, and tests covering concurrency, recovery, migration, path containment, Windows locking, and POSIX directory synchronization.
Observed
- License
- MIT
- Primary runtime
- Python 3.12+ backend with a Node.js 22+ frontend toolchain
- Installation
- Repository clone followed by an interactive Make setup target; Docker and local development are documented
- Interfaces
- Browser application, terminal workbench, embedded Python client, MCP server, and HTTP APIs are documented
- Model connectivity
- Supports LangChain model classes, OpenAI-compatible gateways, vLLM, Codex CLI, and Claude Code OAuth configuration
- Platform support
- Make automation includes Windows handling through cmd.exe and Git Bash, plus non-Windows shell paths
- Testing structure
- Backend test files are documented for memory storage, migration, routing, custom agents, tracing, and model creation
Read from README.md, Makefile, docs/OPENVIKING.md, docs/SKILL_NAME_CONFLICT_FIX.md, docs/CODE_CHANGE_SUMMARY_BY_FILE.md, docs/plans/STORAGE_REWRITE_PLAN.md, docs/plans/STORAGE_REWRITE_CHANGES.md, docs/plans/2026-07-17-remember-login.md, docs/plans/2026-04-01-langfuse-tracing.md, docs/agents/maintainer-orchestrator-design.md, docs/plans/OPENVIKING_HTTP_MEMORY_INTEGRATION.md.
What it can do
Orchestrate multiple AI sub-agents to complete complex tasks
Task definition and requirements → Coordinated results from multiple specialized agents
Execute long-horizon research tasks
Research topic and parameters → Comprehensive research findings and analysis
Generate and execute code for programming tasks
Programming requirements and specifications → Working code and implementation
Handle creative tasks and content generation
Creative brief and requirements → Creative content and deliverables
Manage persistent memory across task sessions
Task data and context information → Stored and retrievable task memory
Execute code in isolated sandbox environments
Code and execution parameters → Safe code execution results
Integrate external tools and services
Tool configurations and API connections → Enhanced capabilities through tool integration
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.