Open-Source Tools Overview: GPT-Researcher, STORM, smolagents...

🌏 中文版

The previous articles covered academic papers. This one covers the open-source world's implementations.

The deep research open-source ecosystem has evolved from "single frameworks" to "tool clusters." Each project solves different problems:

• GPT-Researcher: Emphasizes multi-agent collaboration and report quality
• STORM: Simulates expert conversations to produce structured outlines
• smolagents: Focuses on state management and tool coordination
• Local Deep Research: Fully local deployment
• Feynman: Local-first, can read papers, verify claims

Open-Source Ecosystem Landscape

By Design Philosophy

Detailed Comparison

GPT-Researcher

• GitHub: assafelovic/gpt-researcher
• Core: Autonomous agent researches web + local files, produces long reports with citations
• Features:
  • Intelligent image extraction (scrapes relevant images from web)
  • 2000+ word reports
  • Multi-agent mode (simulates STORM)
  • Deep Research features: Tree exploration mode
  • MCP support
• Strengths: Active community, rich features, reached DeepResearchGym benchmark first place
• Best for: Research tasks requiring rich reports and images

STORM (Stanford)

• Core: Simulates expert conversations to produce research outlines
• Process: First produces outlines via "simulated expert discussion," then retrieves evidence based on outlines
• Features: Highly structured long-form articles, approaching Wikipedia quality
• Strengths: Especially clear structure, coherent narrative
• Best for: Highly structured background articles or overview reports

smolagents

• Core: Lightweight agent framework emphasizing state management
• Features:
  • State as variables (not dependent on LLM memory)
  • Images/audio can be stored as state and reused
  • MCP tool support
• Strengths: Reliable state tracking, suitable for multimodal
• Best for: Long-running tasks with complex state

Local Deep Research

• Core: Fully local deployment, no external API dependencies
• Best for: Privacy-sensitive scenarios, offline environments
• Limitations: Model capabilities and tool coverage restricted

Feynman

• Core: Local-first research assistant
• Capabilities: Read papers + search web + draft + run workflows + verify claims
• Best for: Academic research, paper review, experiment reproduction

Benchmark Performance

According to DeepResearchGym (CMU, May 2025):

GPT-Researcher ranked first in all three categories on DeepResearchGym (>85% citation precision, >80% report clarity, highest key point recall).

Design Philosophy Trade-offs

Trend Observations

1. MCP integration: Most projects are starting to support MCP (Model Context Protocol)
2. Tree exploration: From linear search to tree-branch exploration
3. Multimodal: From text to images, tables, code
4. Local-first: Privacy concerns driving local deployment solutions

References

• GPT-Researcher — Multi-agent deep research.
• STORM — Stanford, expert-conversation outlines.
• smolagents — Hugging Face, state management.
• DeepResearchGym — CMU benchmark.
• Tongyi DeepResearch — order 5 of this series, Tongyi Lab complete system.
• deep-research-survey-overview — Previous article: three-phase landscape classification.