What it does
5 features
DeerFlow · Agent platform
Product and access
Available offering; access methods and prerequisites are described separately in this profile.
Core features
5 features from provider sources · reviewed on 26 October 2026. Availability and limitations follow below.
Usable AI models
Open choice
DeerFlow ships with no fixed model: you enter provider and key yourself, in the config or through presets in the web UI. This works for OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Kimi, MiniMax, GLM and Doubao, for OpenRouter and any OpenAI-compatible endpoint. Ollama, llama.cpp and vLLM run locally, and the Codex and Claude Code CLIs can also be connected.
Model families named
- OpenAI (GPT)
- Anthropic (Claude)
- Google (Gemini)
- DeepSeek
- Alibaba (Qwen)
- Moonshot (Kimi)
- MiniMax
- Z.ai (GLM)
- In-house model
- Other models can be connected freely
- Model choice
- You choose
- Your own API key
- Yes
- Router or custom endpoint
- Yes
- Local models
- Yes
The rating covers the generally available main access. Documented · Sources accessed on
Does it fit my task?
2 practical tasks · 3 requirements and limits
A good fit if …
- you want to run an agent runtime yourself with Docker or locally
- you want to start tasks via Slack, Telegram or Feishu
- you want to trace runs with LangSmith or Langfuse
Less suitable if …
- you want a hosted service with support and a service agreement
- you require documented approval before every write action
Editorial assessment based on provider documentation. Not a product test.
Practical tasks
Editorial examples based on documented features. No product tests performed.
Subagents
- Your input
- Delegate complex tasks to specialized subagents.
- Possible result
- A work product for your review.
- What to check
- Review results and write actions before accepting them.
Sandbox
- Your input
- Run tools in configured execution environments.
- Possible result
- A work product for your review.
- What to check
- Review results and write actions before accepting them.
Delegate complex tasks to specialized subagents.
Your possible prompt
Compare three providers for [task]. Use their official documentation. Show differences, sources, and open questions in a table.
Adjust the task, data and permissions.
Requirements and limits
Limitation
Usage costs
MIT-licensed source code; model usage and infrastructure are funded separately.
Source 1
Can I use it?
With requirements
Europe & regional availability
Consider installation and provider access separately. A complete regional feature matrix is not published.
Reviewed scope: Product access in Germany and the EU
Reviewed on:
With requirements
Access and connected services
Requires your own installation, model access and configured tools.
Region, reason and source
Region covered: Germany / EU
Documented reason: The reviewed sources document no specific restriction caused by the AI Act, DMA or GDPR. Model and service access depend on country and contract.
Accounts & permissions
Based on provider documentation. Permissions depend on the account, integration and organization policies. Review the permissions actually requested before use.
Configuration covered: Self-hosted DeerFlow from the open-source repository
Sources reviewed:
Show all 3 points
Documented with conditions
Data access & team sharing
The local sandbox provider has the shell disabled by default; the Docker provider offers isolated shell access. According to the configuration documentation local mode runs directly on the host and is not a strong isolation boundary. Message channels are Feishu, Slack and Telegram.
Evidence · Evidence
Memory and conversation continuity
5 of 6 criteria documented
Automatically extracted facts with scope checks, manageable entries and backend selection.
Public main branch; pluggable memory backends
Sources accessed:
Documentation by criterion
StorageDocumented
Storage and scope
The backend README describes structured user context, history and scored facts in JSON. Newer backends are pluggable; storage depends on the configured manager.
SelectionDocumented
Selection and timing
Middleware automatically extracts durable user facts and batches updates with a configurable delay. Scope and confidence checks limit promotion.
Edit & deleteDocumented
Inspect, edit and delete
The API reads, creates, updates and deletes facts and the full store. Support for individual management operations depends on the memory backend.
Access boundariesDocumented
Agent and channel boundaries
API access resolves a user identity; automatic global promotion requires durable user facts. Task- or project-specific facts are excluded.
RetrievalDocumented
Retrieval
Documented behavior injects important facts and context into prompts. Pluggable backends may provide search; the selected operating mode determines retrieval.
New conversationsPartly documented
Session reset and compaction
Memory spans conversations; summarization separately configures context summaries. Reset and retention depend on the chosen backend.
Cost and effort
Prices as of 26 October 2026
Prices and included usage
Pricing checked
MIT-licensed source code; model usage and infrastructure are funded separately.
- MIT-licensed source code; model usage and infrastructure are funded separately.
Setup
Requires your own installation, model access and configured tools.
Operation & oversight
Based on public provider documentation. Assessments are editorial inferences from the stated properties. Actual time requirements and reliability have not been measured.
Deployment covered: Self-hosted instance with gateway, Docker or local operation
Sources reviewed:
Editorial assessment
You run the database, keys and tracing yourself. The backend README names no general approval before write actions, so restrict tool permissions and review results yourself.
Show all 6 points
Documented with conditions
Files and working state
Application tables for runs, threads and run events live in SQLite or Postgres. The memory feature extracts facts and preferences from conversations into a JSON file and injects top facts into prompts. The sandbox is per thread.
EvidenceOpen in reviewed sources
Sign-ins
The configuration page names no validity period for sessions or service connections. In the managed model catalogue a missing or wrong key fails closed; other stored values are then treated as missing.
EvidenceDocumented with conditions
Human intervention
An ask_clarification tool lets the agent ask questions, and middleware then interrupts execution. The backend README names no general approval before write actions.
EvidenceDocumented with conditions
Detecting errors
LangSmith and Langfuse can be used for tracing; if a provider is explicitly enabled with invalid credentials, DeerFlow raises an error instead of silently disabling tracing. An endpoint returns debug and audit events per run.
EvidenceDocumented with conditions
Resuming after failures
After a sandbox container ends unexpectedly, the Docker provider drops the dead entry so a thread can get a fresh sandbox. Durable batch tasks keep their lease during initialisation, execution and acceptance checks. The README does not describe resuming interrupted runs.
Evidence
Evidence and open questions
1 source · 1 open point
Open points
The reviewed sources contain no reliable information on these points. A gap says nothing about the product.
Requirements and limits: not yet confirmed
Sources and evidence
Official sources reviewed. Example tasks are editorial suggestions; hands-on testing is outstanding.
Editorial team: Michael Meierhoff · sources reviewed on . Provider information is not independent evidence of performance.
Provider
Open agent runtime for research, coding and multi-step tasks on your own infrastructure.
- Operator
- ByteDance
- Organization
- Provider
- Website
- github.com
Changes to this profile 1
Open agent runtime for research, coding and multi-step tasks on your own infrastructure.
These notes describe editorial changes. Source and pricing review dates appear in their respective sections.