OpenSeed

AI-powered research workflow CLI.
Discover, read, analyze, and synthesize academic papers with Claude — from a single search to a full autonomous research report.

---

Why OpenSeed?

• One command, full report — openseed research run "topic" runs multi-round discovery, analysis, and synthesis autonomously.
• Real ranking — papers ranked by actual citation counts from Semantic Scholar, not keyword matching.
• Claude-native — every analysis step powered by Claude: summarize, review, compare, Q&A, code gen.
• Your library, your tools — MCP server exposes your paper library to Claude Code / Claude Desktop.
• Local-first — SQLite storage, no cloud dependency beyond the LLM API.

---

Quick Start

pip install openseed

openseed doctor    # check environment
openseed setup     # configure auth + model

Auth — any of these work:

export ANTHROPIC_API_KEY=sk-ant-...   # Anthropic API key
claude setup-token                     # OAuth (openseed setup detects it)

---

What You Can Do

Search & Discover

openseed paper search "diffusion models" --count 20
openseed agent search "multi-agent systems"          # deeper search with trend summary

Autonomous Research

openseed research run "ViT image classification"
# discover → analyze → synthesize → markdown report

Manage Your Library

openseed paper add https://arxiv.org/abs/1706.03762
openseed paper list
openseed paper show <id>

Analyze Papers

openseed agent summarize <id>            # structured summary
openseed agent summarize <id> --cn       # Chinese summary
openseed agent review <id>               # peer review
openseed agent compare <id1> <id2>       # side-by-side comparison
openseed agent ask "What is RLHF?"       # research Q&A
openseed agent codegen <id>              # generate experiment code

MCP Server

openseed mcp    # expose library as Claude tools

Available tools: library_stats, search_papers, get_paper, get_graph, search_memories, ask_research

---

How It Works

Search query
     ↓
  ArXiv + Claude WebSearch ── find candidates
     ↓
  Semantic Scholar ────────── rank by citations, fetch metadata
     ↓
  PDF extraction (PyMuPDF) ── full text
     ↓
  Claude analysis ─────────── summarize · review · compare · synthesize
     ↓
  SQLite library ──────────── persist with knowledge graph + FTS5
     ↓
  MCP / Web / CLI ─────────── access from anywhere

---

Architecture

src/openseed/
├── cli/                 Click CLI: paper, agent, research, experiment
├── agent/               AI-powered analysis
│   ├── autoresearch     autonomous multi-round research engine
│   ├── reader           structured summarize / analyze
│   ├── assistant        freeform research Q&A
│   ├── discovery        paper discovery via Claude + Semantic Scholar
│   ├── compare          side-by-side paper comparison
│   ├── strategy         gap analysis + reading recommendations
│   ├── memory           FTS5-backed conversation history
│   ├── context          library-aware prompt assembly
│   └── latex            related-work export with BibTeX
├── services/            External integrations
│   ├── arxiv            ArXiv metadata fetch + search
│   ├── scholar          Semantic Scholar API
│   ├── pdf              PDF text extraction
│   ├── rss · watch      feed discovery + scheduled watches
│   ├── cron · digest    crontab management + digest generation
│   └── sharing          session export/import
├── storage/             SQLite CRUD + knowledge graph + connection pool
├── models/              Pydantic v2 data models
├── mcp/                 MCP server (library as Claude tools)
├── web/                 FastAPI dashboard
├── auth.py              Anthropic client factory
├── config.py            paths, default model
└── doctor.py            environment health checks

---

Star History

---

License

MIT © 2025 cklxx