# Patent MCP Server

> 🚀 **中国专利最准确的开源 MCP。** Give your AI agent the ability to read CN patents with real accuracy — plus global coverage.

[![Tests](https://img.shields.io/badge/tests-32%2F32-brightgreen)](https://github.com/deeparchi-ai/patent-mcp-server/actions)
[![Python](https://img.shields.io/badge/python-3.10%2B-blue)](https://www.python.org/)
[![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
[![MCP](https://img.shields.io/badge/MCP-compatible-purple)](https://modelcontextprotocol.io/)
[![PyPI](https://img.shields.io/badge/pypi-v1.7.0-blue)](https://pypi.org/project/deeparchi-patent-mcp/)

An MCP (Model Context Protocol) server that gives AI agents access to patent data — **CN patents with CPC-aware correction**, plus US/WO global coverage. Runs locally on your machine. No external API, no subscription. Always MIT.

---

## Why Self-Deployed

- **It's just Python.** Install it, your agent uses it. No server to maintain, no credential to share.
- **No API key for 80% of use cases.** Patent details and claims come straight from Google Patents public pages.
- **Your data stays local.** Nothing leaves your machine except the same HTTP requests a browser would make.
- **BigQuery search is optional.** Only turn it on if you need full-text search across 1.4B records.

---

## 30-Second Install

```bash
pip install deeparchi-patent-mcp
```

Or from source:
```bash
git clone https://github.com/deeparchi-ai/patent-mcp-server.git
cd patent-mcp-server
pip install -e .
```

---

## Quick Start

Pick your agent platform and add this to its MCP config:

### Claude Desktop

```json
{
  "mcpServers": {
    "patent-mcp": {
      "command": "python",
      "args": ["-m", "src.server"],
      "cwd": "/path/to/patent-mcp-server"
    }
  }
}
```

### Cursor / Windsurf / Cline

Same config as Claude Desktop above.

### Hermes Agent

```yaml
mcp_servers:
  patent-mcp:
    command: "python"
    args: ["-m", "src.server"]
    workdir: "/path/to/patent-mcp-server"
```

---

Now ask your agent:

> *"Get patent US-7650331-B1 and summarize the claims."*

---

## What's Included

| Tool | What It Does | Needs Setup? |
|------|-------------|:---:|
| `get_patent` | Full patent details: classifications, citations (X/Y/A/D), inventors, assignees, family | No |
| `get_patent_claims` | Patent claims text — legal scope. Supports US, CN, and most countries via Google Patents | No |
| `search_patents` | Search 1.4B patents by keyword, CPC, assignee, country, date range | Optional GCP |

The first two cover 80% of use cases. Zero cost. Zero setup.

### CN Patent Search (v1.7.0)

Three-layer discovery for Chinese patents:

| Layer | Backend | Cost | When |
|-------|---------|:---:|------|
| 1. BigQuery | Google Patents Public Data | Free tier | `cpc=H01L + country=CN` |
| 2. Firecrawl | Web search fallback | 4 credits/query | BigQuery cost-rejects specific CPC (e.g. H01L25/065) |
| 3. Google Patents | Detail enrichment via proxy | Free | All patent detail lookups |

- **Keyword search works:** `query="芯片" + country=CN` searches both English AND Chinese abstracts (v1.5.2 fix).
- **Assignee filter:** `assignee="BOE" + country=CN` → company/city-level patent landscape.
- **CPC classification:** Use parent CPC classes (`H01L`) for broader CN coverage; specific CPC codes (`H01L25/065`) trigger web fallback.
- See [`SEARCH_GUIDE.md`](SEARCH_GUIDE.md) for detailed search strategy and tested CPC codes.

---

## Optional: Enable BigQuery Search

If you need `search_patents`, add a GCP project:

1. Create a [GCP project](https://console.cloud.google.com/) with BigQuery enabled
2. Create a service account, download JSON key
3. Set env vars:
   ```bash
   export GOOGLE_APPLICATION_CREDENTIALS="/path/to/key.json"
   export GCP_PROJECT_ID="your-project-id"
   ```
4. Copy the wrapper template and fill in your paths:
   ```bash
   cp run.sh.example run.sh
   # Edit run.sh → set your GCP paths
   ```

BigQuery free tier: 1 TB/month — individual use is essentially free.

---

## Advanced: Team Server (HTTP/SSE)

Need multiple people to share one patent-mcp instance? Start it as an HTTP server:

```bash
cp run-http.sh.example run-http.sh
# Edit → set GCP creds (skip if only using web tools)
PORT=8090 ./run-http.sh
```

Team members connect with:
```yaml
mcp_servers:
  patent-mcp:
    url: "http://<server-ip>:8090/sse"
```

A [systemd service template](run-http.sh.example) is included for production deployment.

---

## How It Works

```
┌──────────────┐     ┌───────────────────────────────────────┐
│  AI Agent    │────▶│  patent-mcp-server                    │
│  (Claude,    │     │  (runs on YOUR machine)               │
│   Cursor,    │     │                                       │
│   Hermes)    │     │  search_patents:                      │
│              │     │    ┌──────────┐    ┌───────────────┐  │
│              │     │    │ BigQuery │───▶│ Firecrawl     │  │
│              │     │    │ (primary)│    │ (CN fallback) │  │
│              │     │    └──────────┘    └───────┬───────┘  │
│              │     │                           │           │
│              │     │  get_patent / get_patent_claims:      │
│              │     │    ┌──────────────────────┐           │
│              │     │    │ Google Patents (web) │           │
│              │     │    │ → BigQuery fallback  │           │
│              │     │    └──────────────────────┘           │
└──────────────┘     └───────────────────────────────────────┘
```

- **Web scraping for details** — fast (~1.5s), free, no credentials
- **BigQuery for search** — 1.4B records, CN full-text, optional
- **Firecrawl for CN fallback** — kicks in when BigQuery cost-rejects specific CPC queries
- **Smart fallback** — every tool tries web first, auto-falls to BigQuery if you have it

---

## Tools Reference

### `get_patent`

```
get_patent(publication_number="US-7650331-B1")
```

Returns: classifications, citations (X/Y/A/D prior art markers), family ID, dates, inventors, assignees. Cites prior art markers so your agent can assess novelty at a glance.

**CN patent note:** Google Patents web scraping provides machine-translated English data for CN patents. CPC codes from web scraping are empty (JS-rendered). For CPC, use BigQuery path.

### `get_patent_claims`

```
get_patent_claims(publication_number="US-7650331-B1")
```

Returns: full claims text. Supports US, CN (machine-translated English), and most countries via Google Patents web scraping.

### `search_patents`

```
search_patents(cpc="G06N", country="CN", after="2023-01-01", limit=5)
search_patents(assignee="TSMC", country="CN")
search_patents(query="芯片", country="CN")          # keyword search (CN: searches both EN+ZH)
```

Search 1.4B patents by keyword, CPC classification, assignee, country, date range. For CN patents, keyword search scans both English and Chinese abstracts.

**Cost control:** Queries require at least one filter (`cpc`/`country`/`assignee`/`after`). A dry-run budget guard rejects queries over 50 GB. When BigQuery rejects a CN CPC query (e.g., `H01L25/065` → 256 GB), the web fallback automatically searches via Firecrawl.

See [`SEARCH_GUIDE.md`](SEARCH_GUIDE.md) for best practices and [`docs/cn-cpc-correction-table.md`](docs/cn-cpc-correction-table.md) for tested CPC codes.

---

## Development

```bash
pip install -e ".[dev]"

pytest tests/ -v          # 32 tests, ~1.5s
ruff check src/ tests/    # lint
mypy src/                 # type check
```

---

## License

MIT — see [LICENSE](LICENSE).

---

## Author

[DeepArchi OPC](https://github.com/deeparchi-ai) — AI agent infrastructure for enterprise architecture.