DeerFlow

DeerFlow (Deep Exploration and Efficient Research Flow) is a super-agent harness: a runtime that gives an agent a filesystem, memory, sandboxed execution, and the ability to plan work and spawn sub-agents. Version 2.0 is a ground-up rewrite that shares no code with version 1, which continues as a Deep Research framework on a separate branch. The 2.0 runtime is built on LangGraph and LangChain and is meant to be used as shipped or modified.

Capabilities arrive as skills: Markdown modules that define a workflow plus supporting resources. Built-in skills cover research, report generation, slide creation, web pages, and image and video generation, and you can add, replace, or combine them. Skills load progressively, only when a task needs them, so the context window stays smaller; you can force one for a single turn by starting a request with /skill-name. A skill can declare allowed-tools, which filters both the tool schemas the model sees and what it may execute once that skill is active.

Each task runs against its own execution environment with a filesystem view covering skills, workspace, uploads, and outputs. Providers include a local sandbox, Docker or the all-in-one image, a Kubernetes provisioner, and E2B microVMs; the local provider is a managed tool-path boundary rather than host isolation, so the container and VM providers are the ones that keep the filesystem boundary enforceable alongside shell access. After each run DeerFlow records a workspace change summary and shows created, modified, and deleted files with text diffs.

Sub-agents are treated as an optimization rather than the default: the lead agent spawns them when parallel latency, specialist capability, or context isolation gives a clear benefit, then verifies and synthesizes their structured results. Thread-level controls include /goal to attach one completion condition that survives across turns, and /compact to summarize older context while the full chat stays visible.

DeerFlow is model-agnostic and works with any OpenAI-compatible endpoint, with configuration examples for hosted APIs, OpenRouter-style gateways, self-hosted vLLM, and CLI-backed providers that reuse Codex CLI or Claude Code credentials. It runs from a web UI, a terminal workbench, IM channels, an embedded Python client, and an MCP server, and can emit traces to LangSmith, Langfuse, or Monocle.

Features

• Skills system: Markdown capability modules loaded on demand, with slash activation and per-skill tool policies
• Sub-agent delegation: scoped contexts, termination conditions, structured results, and bounded runtime capacity
• Sandboxes: local, Docker/all-in-one, Kubernetes provisioner, and E2B providers with per-task filesystems
• Session goals: /goal attaches one completion condition that a typed evaluator checks after each run
• Context compaction: /compact summarizes older turns while the visible transcript is preserved
• Long-term memory: persistent state across sessions alongside per-run working files
• Scheduled tasks: recurring runs, plus IM channel delivery of results
• Model configuration: any OpenAI-compatible endpoint, vLLM reasoning toggles, and CLI-backed provider entries
• Tracing: LangSmith, Langfuse, and Monocle integrations for run inspection
• Interfaces: web UI, terminal workbench, embedded Python client, MCP server, and IM channels