Open-source · MCP · Apache-2.0

Patent and trademark data, built for AI agents.

A unified Python library and MCP server over USPTO, EPO, public patent search, the MPEP, the TMEP, and CPC. Sixty-three typed tools your agent can call directly — with no seat license, no contract, and no rate-limited vendor scrape in front of the official data.

Try the hosted demo → Install locally View on GitHub →

63 typed MCP tools across USPTO, EPO, and public patent search

12 upstream sources unified behind one async API

$0 seat licenses. Public data, open source, no contract.

What it does

Ask your agent the questions you'd ask a research analyst.

Six real prompts. Each one fans out to between two and a dozen MCP tool calls under the hood — search, status lookup, document fetch, family resolution — without you wiring any of it.

Find the most-cited recent patents on transformer inference acceleration and summarize the key claims.

What's the full prosecution history for US Patent 11,234,567?

Map the patent landscape for mobile display technology over the last five years.

Track the legal status of EP3456789 across every designated state.

Search the TMEP for guidance on Section 2(d) likelihood-of-confusion refusals.

What's the current status of trademark serial 97123456, and who filed it?

Why it exists

The missing layer between the APIs and your agent.

Patent data lives in three places today. Official free APIs from the USPTO and EPO are powerful but raw — REST surfaces with no SDK, no caching, no agent affordances. Enterprise IP analytics platforms run five-to-six figures a year on seat-based contracts, with UI-only access and no programmable layer. The new wave of AI patent products wraps closed dashboards around the same data, sold to attorneys, with no API to call.

None of those are built for an agent. patent-client-agents is the layer in between — a unified, async, MCP-first interface over the free public sources, designed to be called by code (and by AI agents calling code).

What requests is for HTTP, patent-client-agents aims to be for patent data: a small, well-typed, batteries-included client you'd reach for without thinking.

Coverage

Twelve sources, one API.

Patents, trademarks, examination history, classifications, and the official examiner manuals — across the USPTO, the EPO, and the world's public patent indexes.

Source	What you get
Public global search	Worldwide patent index — full text, citations, PDFs, families
USPTO ODP	Applications, prosecution history, PTAB trials & appeals, petitions, bulk data
USPTO Public Search	PPUBS full-text search and document retrieval
USPTO Assignments	Patent ownership transfers and reel/frame lookups
USPTO Office Actions	Rejection analytics, cited references, full OA text
USPTO TSDR trademarks	Status, prosecution documents, mark images, batch lookup
USPTO TM Assignments trademarks	Trademark ownership transfers from the Assignment Center
EPO OPS	European patents, Inpadoc families, legal events, EP Register
MPEP	Manual of Patent Examining Procedure — search and section lookup
TMEP trademarks	Trademark Manual of Examining Procedure — search and section lookup
CPC	Classification hierarchy lookup, search, and CPC/IPC mapping
JPO library only	Japanese patents, examination history, PCT national phase

Every source ships with automatic caching, rate limiting, and retry logic — so an agent that calls the same patent twice pays the network cost once.

Three ways to use it

Same code underneath. Pick your surface.

The hosted MCP demo for the fastest path, a Claude Code plugin for local power use, or the raw Python library for your own apps.

For AI agents

Hosted MCP demo

Point any MCP-speaking client at the public endpoint. Sign in once with a verified account. No tokens, no setup. Rate-limited to 100 MB / day per account.

{
  "mcpServers": {
    "patent-client-agents": {
      "url": "https://mcp.patentclient.com/mcp"
    }
  }
}

Open the demo

For Claude Code users

Plugin install

Three slash commands inside Claude Code adds 49 patent and trademark tools (61 with JPO credentials). Server runs locally under uvx.

/plugin marketplace add parkerhancock/patent-client-agents
/plugin install patent-client-agents@patent-client-agents
/reload-plugins

Plugin install guide

For Python developers

Async library

Import the clients directly. Pydantic v2 models, async context managers, hishel caching, tenacity retries. No MCP runtime required.

pip install patent-client-agents

async with UsptoOdpClient() as c:
    app = await c.applications.get("17/000,000")
    print(app.title)

Library quickstart

How it fits together

One thin server. One typed library. Twelve upstream sources.

The MCP server is a translation layer over an async Python library. Use either independently — agents through MCP, humans through Python — and reach the same data.

flowchart TB
  A["Your AI Agent"]
  M["patent-client-agents MCP Server"]
  L["patent_client_agents Python library"]
  U["USPTO patents
ODP · PPUBS"]
  T["USPTO marks
TSDR · TM assignments"]
  E["EPO OPS · CPC"]
  G["Public patent search
MPEP · TMEP"]
  J["JPO ★ library only"]

  A --> M
  M --> L
  L --> U
  L --> T
  L --> E
  L --> G
  L --> J

  classDef agent fill:#0d1530,stroke:#0d1530,color:#ffffff
  classDef server fill:#eaf0fb,stroke:#2a5fdf,color:#0d1530,stroke-width:1.5px
  classDef lib fill:#2a5fdf,stroke:#2a5fdf,color:#ffffff
  classDef source fill:#ffffff,stroke:#d4dbe6,color:#0d1530
  classDef muted fill:#f7f9fc,stroke:#d4dbe6,color:#5b6781,stroke-dasharray:3 3

  class A agent
  class M server
  class L lib
  class U,T,E,G source
  class J muted

★ JPO is library-only — JPO MCP tools are not available.

Built right

The way you'd build it yourself.

Open source, well-tested, async-first. Nothing locked away in a SaaS dashboard.

Apache-2.0 Use it commercially, fork it, run it on your own infrastructure. No surprise rug-pulls.
1,056 tests on VCR cassettes Deterministic. No live API hits in CI. Replay every cassette in fifteen seconds.
Shared HTTP scaffolding One BaseAsyncClient, one cache (hishel + SQLite), one retry policy (tenacity), one exception hierarchy.
Self-hostable MCP server Terraform and Dockerfile included if you want a private instance with your own auth on any container platform.
Standing on patent_client The original library that this project succeeds. Eight years of accumulated quirks already fixed.
Pydantic v2 models Typed responses, IDE autocomplete, no dict-fishing. Errors at parse time, not at runtime.