Patent Connector 1.6: the CLI lands, trademarks start talking, and the data gets deeper

Wolfgang Stark

Wolfgang Stark

- 4 min read
Share
Patent Connector 1.6: the CLI lands, trademarks start talking, and the data gets deeper

Patent Connector 1.6 is the biggest release since launch. It does three things at once: it ships the long-promised agentic-AI-first command-line tool, it teaches the connector to handle sound and multimedia trademarks, and it goes back through every office already supported and pulls more out of each one. No scraping anywhere - every patent, trademark and design field comes straight from an official office API, queried with real credentials. There is no intermediate database in between, which matters for more than freshness: you get the office's most current data, and nothing sits in the middle to fall behind, drop fields, or quietly corrupt them. The failure mode of an aggregator - whether it is a free re-index like Google Patents or a paid commercial database - is not only delay, it is data quality.

Here is what is new.

The patent CLI - built agentic-AI-first

The MCP server puts patent tools inside a chat. The CLI is for the next step: agents. It is built agentic-AI-first - a single self-contained binary that an autonomous agent can shell out to and get official, structured patent data back, the same tools the chatbot has, with JSON (or YAML, or a table) on stdout that a model can parse directly. Coding agents, research agents, and your own in-house agents get first-hand IP data without an SDK, a wrapper, or a browser.

brew install patent-dev/tap/patent
patent login
patent search-patents --query "solid state battery" --limit 5 -o table

Humans scripting pipelines and CI get the exact same surface, of course - but the design target was the agent in the loop, not just the shell prompt. And it runs three ways: against the hosted server (zero setup, free tier), against your own patent server for a team with shared credentials, or fully in-process with your own office API keys. One command, three places it can execute: Patent CLI

Trademarks that talk: sound, video, and 3D marks

Trademarks are not only words and logos. The EU register holds sound marks, multimedia marks, and 3D-model marks - and until now there was no good way to get at them from an AI tool. Patent Connector 1.6 returns them as a playable link: ask for a mark's sound, get a URL you can open and actually hear.

patent get-trademark-image --serial-number 017396102 --provider euipo --media-type sound

The same mechanism covers video marks and 3D models, and on the design side too. Images still come back as images; the new part is that the non-visual marks are no longer a dead end.

A personal aside: I love anything with sound and music, so making trademarks actually playable - turning a sound mark from a dead reference into something you click and hear - was easily the most fun part of this release to build. A close second was the architecture underneath it: an MCP server, an MCP client, and a standalone CLI living in a single binary behind one clean separation of concerns. That part earns its own write-up, which I will get to soon.

More from every office

The quieter half of 1.6 is depth. Rather than only adding new jurisdictions, this release went through the existing ones and surfaced data that was already in the APIs but not yet exposed:

  • United States (USPTO): bibliography now carries the inventors, the agent or firm of record, CPC classifications, and foreign-priority claims (with the office name normalized to an ISO code). Assignments now include the assignee's location and a direct document download link. Patent search gained a server-side filing-date range. And two new families of tools: petition decisions and the full PTAB surface (trials, appeals, interferences, and their decision documents).
  • European Union (EUIPO): search the Nice (trademark goods and services) and Locarno (design product indications) harmonized term databases, and owner / representative search to pull a party's whole portfolio.
  • Germany (DPMA): full bibliography - all classifications and versions, priorities, and every party - plus the IP-right type (patent, utility model, or SPC), and downloadable publication PDFs.
  • Japan (JPO): application-number cross-referencing, priority-basis family, patent term, and J-PlatPat links. Prosecution document bundles are now parsed into individual notices - examiner, date, and decoded body text - instead of an opaque ZIP.
  • Australia (IP Australia): trademark and design images, search sorting and status filters, and representatives plus earliest priority date in the bibliography.
  • Europe (EPO): CPC classifications in the bibliography and proper search paging.

Jurisdiction-neutral term discovery

This one is an update to the reference server itself. Reference search - the built-in, no-credentials index over classification schemes and examination manuals - now also covers the Nice (trademark goods and services) and Locarno (design) term classifications, and it runs them the same way it runs CPC, IPC and the manuals: hybrid search, fusing exact keyword matching with semantic vector search over an embedded index. Ask "what Nice class is milk?" or "find the Locarno term for a chair" and get the harmonized terms back, grouped so jurisdictions are never co-ranked.

patent reference-search --domain nice --query "e-bike"
patent reference-search --domain locarno --query "robot vacuum dock"

Where the hybrid part earns its keep is when your words are not the register's words. Search Nice for "e-bike"and the top hit is "electric bicycles" - the harmonized term you actually need, which shares no whole word with your query, so a keyword search would miss it while the vector side recognizes the concept. The same on the design side: a Locarno search for "robot vacuum dock" surfaces "Charging stations for robotic vacuum cleaners" - a dock is a charging station, obvious to a human and to the embeddings, invisible to a literal match. In both cases the keyword half keeps any exact hits honestly ranked; the semantic half catches the paraphrase. (Both examples are real - run them yourself.)

Taiwan out of beta, and two more offices on the horizon

Taiwan (TIPO) graduates out of beta in this release - param-based patent search by applicant, classification, and date range is live and validated against the production endpoint.

And the dashboard now lists two offices as coming soon: INPI France and KIPO / KIPRIS (Korea). They are announced, not yet live - if you already have API access to either and want in early, there is a "Get Access" contact on each card.

Available now

Patent Connector 1.6 is live. In the chatbot, reconnect your MCP client to pick up the new tools. For the CLI, brew install patent-dev/tap/patent (or the curl installer / Scoop). New here? Sign up at patent.dev/patent-connector.

Questions about integrating patent, trademark, or design data into your AI workflows - or running it inside your own infrastructure? Reach out through patent.dev.

Read more