Research Guard

Check your research idea's novelty before investing months into it.

CI Python 3.11+ License MCP

What it does

Research Guard is an MCP server that tells you if your research idea has already been explored. It decomposes your idea into (Problem, Method, Innovation), searches 200M+ academic papers across Semantic Scholar and OpenAlex, and gives you a RED / YELLOW / GREEN verdict with evidence.

Not a search tool. A research decision guard.

Quick start

1. Install

uvx research-guard-mcp

2. Get API keys

3. Configure your MCP client

{
  "mcpServers": {
    "research-guard": {
      "command": "uvx",
      "args": ["research-guard-mcp"],
      "env": {
        "S2_API_KEY": "your-semanticscholar-api-key",
        "OPENAI_API_KEY": "your-openai-api-key"
      }
    }
  }
}

claude mcp add research-guard -- uvx research-guard-mcp

{
  "mcpServers": {
    "research-guard": {
      "command": "uvx",
      "args": ["research-guard-mcp"],
      "env": {
        "S2_API_KEY": "your-semanticscholar-api-key",
        "OPENAI_API_KEY": "your-openai-api-key"
      }
    }
  }
}

How it works

Your research idea
       |
   [Decompose]       →  (Problem, Method, Innovation) triple
       |
   [Expand]          →  5-dimension concept expansion
       |                  (synonyms, ancestors, alternatives, critiques, cross-domain)
       |
   [Retrieve]        →  Semantic Scholar + OpenAlex (parallel, deduplicated)
       |
   [Quick Rank]      →  Contribution-point scoring (text + semantic similarity)
       |
   [Deep Compare]    →  LLM pairwise comparison with evidence grounding
       |
   [Assess]          →  RED / YELLOW / GREEN  +  overlaps  +  blind spots

Tools

Example

You: I'm thinking of using vision transformers with shifted windows
     for 3D medical image segmentation. Is this novel?

Claude: [calls check_novelty on Research Guard]

Research Guard result:
  Rating: YELLOW (partially overlapping)
  - Swin UNETR (2022) already uses Swin Transformer for brain tumor segmentation
  - TransUNet (2021) combines transformers with U-Net for medical images
  - Differentiation space: your 3D shifted window approach differs from their
    2D slice-based methods
  - Blind spot: no counter-evidence found for this combination

Claude: Your idea has significant overlap with existing work, but there's
        a clear differentiation path — the 3D shifted window mechanism is
        novel in this domain. I'd recommend proceeding with explicit
        positioning against Swin UNETR.

Configuration

All settings are configured via environment variables (prefix RESEARCH_GUARD_):

Architecture

research-guard/
├── packages/
│   ├── guard_core/          # Core engine (no MCP dependency)
│   │   ├── models.py        # Pydantic v2 type system
│   │   ├── decomposer.py    # LLM-based idea decomposition
│   │   ├── expander.py      # 5-dim concept expansion
│   │   ├── retriever.py     # Multi-source parallel retrieval
│   │   ├── novelty_checker.py # Main pipeline orchestrator
│   │   ├── similarity.py    # Text + semantic scoring
│   │   ├── blind_spot.py    # Counter-evidence detection
│   │   ├── contribution_comparator.py  # Contribution-level comparison
│   │   ├── cache.py         # SQLite async cache
│   │   ├── cost_tracker.py  # LLM cost monitoring
│   │   └── sources/         # API clients (S2, OpenAlex, arXiv)
│   ├── guard_mcp/           # FastMCP server (thin adapter)
│   └── guard_web/           # [Phase 2b] Web UI
├── plugins/                 # Extensible data source ABCs
├── datasets/                # Domain vocabularies + validators
├── tests/
│   ├── unit/                # Pure logic tests
│   └── integration/         # Real API tests (skippable)
└── docs/

Development

git clone https://github.com/Fengrru/research-guard.git
cd research-guard
uv sync

# Run all tests
uv run pytest tests/unit -v

# Lint
uv run ruff check packages/ plugins/ datasets/ tests/

# Type check
uv run mypy packages/ --ignore-missing-imports

Disclaimer

This tool only covers indexed public literature (Semantic Scholar, OpenAlex, arXiv). Unpublished work, gray literature, preprints not yet indexed, and non-English papers may not be fully captured. Always use this as a decision-aid, not an absolute verdict.

License

Apache-2.0 — commercial-friendly, patent grant included.

Contributing

See CONTRIBUTING.md for guidelines. We welcome contributions for:

• New data source plugins (e.g., PubMed, DBLP, IEEE Xplore)
• Domain-specific expansion vocabularies
• Web UI (Phase 2b)