An open-source alternative to Google Gemini Deep Research. Uses a 4-role Agent collaborative architecture that simulates how a real consulting team works, automatically generating in-depth research reports with academic citation footnotes.
- 4-Role Architecture: Agent-manager (planning + QA + delivery) + Agent-researcher (search + data) + Agent-report consultant (writing + visualization self-check) + Agent-data reviewer (independent data verification)
- 5 Decision Gates: User and Agent-manager can send work back upstream at key checkpoints to ensure quality. Plus mandatory visualization compliance check before output
- Visualization Gate: Prevents text-only reports — forces trend data into charts, hierarchies into diagrams, timelines into structured tables. Pandoc mechanical conversion explicitly forbidden
- Source Priority Pyramid: Annual reports/broker research > government data > authoritative media > corporate websites, with strict blacklist filtering
- Anti-Hallucination: research_data.json persistence + Agent-data reviewer independent review + scripts forbidden from hardcoding numbers
- Dual Format Output: .docx (with citation superscripts + bibliography) + .pptx (thesis-statement-driven consulting-grade slides)
- Consulting Framework Library: IBM Five-Forces by default, supports PEST/Porter's Five Forces/SWOT/3C/BCG Matrix
- Completely Free: Uses DuckDuckGo search, no API key required, unlimited usage
| Dimension | Gemini Deep Research | This Skill |
|---|---|---|
| Execution Model | Single submission | Multi-Agent collaboration + human intervention |
| Data Verification | Relies on search ranking | Independent agent dedicated to data review |
| Quality Control | One-shot output | 4 decision gates, supports multiple rounds of revision |
| Anti-Hallucination | Relies on model capability | JSON persistence + independent review + no hardcoded numbers |
| Free Tier | ~5-6 per month | Unlimited |
git clone https://github.com/yslicn/multi-agent-deep-research.git
cd multi-agent-deep-research
cp -r . ~/.claude/skills/deep-researchThis is a prompt-based skill — SKILL.md contains the complete consulting workflow. Any AI agent can follow it.
| Platform | How to Load |
|---|---|
| Claude Code | cp -r . ~/.claude/skills/deep-research/ then type /deep-research |
| OpenCode | Add SKILL.md as a project rule or skill file |
| Codex | Load SKILL.md via cc-switch as a project instruction |
| Hermes | Import SKILL.md as a skill definition |
| Any Agent | Paste SKILL.md content into system prompt or custom instructions |
Help me research the China pre-made meal market using the Five-Forces framework
See full-length reports generated by this system — with structured sections, data tables, charts, and academic-grade citations.
The system employs 4 specialized AI agents that work together in a structured pipeline — each agent has a distinct role, and data flows sequentially from one to the next with built-in feedback loops for quality assurance.
- User submits research request
- Agent-manager analyzes and proposes research approach → Gate 1: User approves?
- Agent-manager breaks down tasks and assigns to agents
- Agent-researcher executes web search, collects data
- Agent-report consultant integrates data, writes draft report
- Visualization self-check: trend data → charts, hierarchies → diagrams, timelines → tables → Gate 2
- Agent-data reviewer independently verifies all cited data → Gate 3: Data trustworthy?
- Agent-manager evaluates report quality (incl. visualization compliance) → Gate 4: Quality meets standards?
- Agent-manager delivers report with push notification
- User receives report → Gate 5: Requirements satisfied?
Automatically generated with academic-grade citation superscripts, structured sections, data tables, and charts. View the full report online for the complete reading experience.
| Framework | Analysis Dimensions | Best For |
|---|---|---|
| IBM Five-Forces (default) | Macro, Industry, Customer, Competitor, Self | Comprehensive industry research, strategy |
| PEST | Political, Economic, Social, Technological | Macro environment analysis |
| Porter's Five Forces | Supplier power, Buyer power, New entrants, Substitutes, Rivalry | Industry structure analysis |
| SWOT | Strengths, Weaknesses, Opportunities, Threats | Company/product diagnosis |
| 3C Model | Company, Customer, Competitor | Competitive positioning |
| BCG Matrix | Market growth vs Relative market share | Product portfolio analysis |
Each research project produces:
workspace/[project-name]/
research_data.json # Persistent data store (single source of truth)
report.md # Full report in Markdown
report.docx # Formatted Word report with citations
review_report.json # Data reviewer verification report
- Every key data point must have a source citation footnote [n]
- Each data point must be verified by at least 2 independent sources
- Data credibility is graded: high / medium / low
- Report length: 8,000–15,000 words (standard), ≥12,000 (deep)
- Reports must include charts (tables, bar charts, line charts)
- Agent-data reviewer must independently verify all data (cannot reuse Agent-researcher's search results)
The following sources are strictly prohibited:
- Baidu Baike, Sogou Baike, 360 Baike (no editorial review)
- Baidu Baijiahao and other unverified self-media
- Weibo, X (Twitter), and other social media
- Content without date, author, or source attribution
- Personal blogs and unverified forum comments
- Relies on publicly available web information; cannot access paywalled databases (Euromonitor, Wind, IBISWorld)
- Data timeliness depends on search tool accessibility
- For highly specialized domains, human review is recommended
- DuckDuckGo search breadth is lower than Google; supplement with other channels when needed
- Agent-data reviewer's independent verification is also limited to public sources
- Model-calculated estimates should be treated as reference, not primary data
- The quality of pptx format is poor by now, still working on it
MIT