Skip to main content
Glama
caiuschrom-a11y

mcp-patent

README.md
# mcp-patent

USPTO patent prior-art search — both an MCP server (Claude/Cursor/Windsurf)
and a standalone CLI. Sister product to mcp-trademark.

## Why this exists

Patent prior-art search has the same shape as trademark watching:
enterprise tools (PatBase, Derwent, Questel) cost $5K-50K/yr. Boutique
firms and indie inventors get squeezed.

Same playbook:
- $19/mo per portfolio (managed)
- $149 lifetime self-host bundle
- $499 lifetime Pro (semantic similarity via Qwen, full claims search)

## What it does

- **Keyword search** across title + abstract + claims (USPTO public API)
- **Prior-art ranking** — score by title overlap (Levenshtein), abstract
  cosine similarity, and CPC class overlap
- **Date filter** — only filings before your priority date
- **MCP tools** — call from Claude/Cursor/Windsurf
- **CLI** — `patentwatch search` / `patentwatch prior-art --title ...`

## Pricing tiers

| Plan | Price | Features |
|------|-------|----------|
| Lifetime self-host | $149 | Full source, MCP server, CLI |
| Lifetime Pro | $499 | + Qwen semantic similarity, claims-text search |
| Managed | $19/mo per portfolio | We host, daily watch, email digest |

## Run locally

```powershell
cd C:\openclaw-products\mcp-patent
python -m venv .venv
.\.venv\Scripts\activate
pip install -e ".[dev]"

patentwatch search "battery thermal management"
patentwatch prior-art --title "Battery cooling system" \
                       --abstract "Liquid-cooled..." \
                       --classes "H01M,F28D" \
                       --before "2026-01-01"
```

## MCP usage (Claude Desktop)

```json
{
  "mcpServers": {
    "patent": {
      "command": "python",
      "args": ["-m", "src.server"],
      "cwd": "C:\\openclaw-products\\mcp-patent"
    }
  }
}
```

## Roadmap
- [ ] Bulk-feed integration for completeness
- [ ] Qwen semantic-similarity scoring (paid Pro tier feature)
- [ ] Claims-text full search
- [ ] EPO + WIPO global patent
- [ ] Image-based prior art (CAD diagram embeddings)