Multi-Agent Automated Research System — an end-to-end research pipeline from a vague idea to a paper draft, powered by four AI Agents.
中文 | English
Prerequisites: Python 3.10+
cd backend
pip install -r requirements.txt
python -m uvicorn main:asgi_app --host 0.0.0.0 --port 3001 --loop asyncio --http h11Or use the launch script at the project root: ./start.sh (macOS/Linux) or start.bat (Windows).
Open http://localhost:3001 to access the homepage.
A Research is the top-level work unit. Each Research corresponds to a research topic and progresses through four stages: refine, plan, execute, and paper.
Enter prompt on homepage -> Create Research -> Enter detail page -> Run full pipeline or stage-by-stage
| Page | Description |
|---|---|
Home (/ or index.html) |
Enter a research topic, click Create to start a Research |
Research List (research.html) |
View all Researches, click to enter details |
Research Detail (research_detail.html?researchId=xxx) |
Execute the four stages: refine / plan / execute / paper |
| Stage | Purpose |
|---|---|
| Refine | Extract keywords, search arXiv, generate refined idea |
| Plan | Decompose the idea into an executable task tree |
| Execute | Run atomic tasks in parallel, validate outputs |
| Paper | Generate a paper draft from Plan and Task outputs (Markdown/LaTeX) |
Each stage supports Run (start fresh), Resume (continue from stopped/failed), Retry (clear and restart), and Stop (abort).
The Thinking panel shows reasoning traces; the Output panel shows final artifacts (literature, task outputs, paper).
| Agent | Responsibility |
|---|---|
| Idea | Keyword extraction, arXiv search, Refined Idea generation |
| Plan | Task decomposition (atomicity -> decompose -> format -> quality) |
| Task | Atomic task execution and validation |
| Paper | Paper draft generation (Markdown/LaTeX) |
Each Agent supports three modes: Mock (simulated), LLM (single-turn), and Agent (tool-use loop). Switch modes in Settings -> AI Config. The Paper Agent's Agent mode is currently an MVP: outline -> sections -> assembly.
Alt+Shift+S (Win/Linux) or Cmd+Shift+S (Mac) to open Settings:
- Theme — Light / Dark / Black
- AI Config — Agent mode, Idea RAG, Self-Reflection, API Preset
- Data — Restore recent plan, Clear all data
maars/
├── backend/ # FastAPI + SSE realtime event bridge
│ ├── api/ # Routes (idea, plan, execution, paper, research, session, settings, ...)
│ ├── idea_agent/ # Idea Agent
│ ├── plan_agent/ # Plan Agent
│ ├── task_agent/ # Task Agent (ExecutionRunner + 5 function modules + DI)
│ ├── paper_agent/ # Paper Agent
│ ├── validate_agent/# Step-B contract review (Task Agent sub-component)
│ ├── shared/ # LLM client, constants, reflection, utilities
│ ├── visualization/ # Execution graph layout computation
│ └── db/ # SQLite persistence (Research metadata + sandbox)
├── frontend/ # Static pages, SSE/EventSource, vanilla JS
│ ├── index.html # Homepage (create Research)
│ ├── research.html # Research list
│ └── research_detail.html # Research detail (four-stage execution)
├── ARCHITECTURE.md # Detailed architecture documentation
└── start.sh / start.bat # Launch scripts
| Document | Description |
|---|---|
| Architecture | System architecture overview (中文) |
| docs/README.md | Documentation index |
| docs/DEVELOPMENT_GUIDE.md | Development guide (architecture, Research API, Skill system) |
| docs/workflow/ | Workflow docs (user flow, Research pipeline, four Agents, realtime events) |
| docs/FRONTEND_SCRIPTS.md | Frontend scripts and module dependencies |
| docs/RELEASE_NOTE_STANDARD.md | Release note writing standard |