TMrex 🦖

TMrex is a Claude skill that runs a screening-level U.S. trademark clearance: it searches a local copy of the USPTO register, the web for unregistered (common-law) use, and domain names, then writes a report covering likelihood of confusion (the DuPont factors), distinctiveness and Section 2 refusal risks, and class coverage.

It is a screening tool, not legal advice. Reports say what was and wasn't searched, and should be reviewed by a trademark attorney.

How it works

• The register. data/uspto.db is a SQLite copy of every USPTO trademark case file (about 14 million), with full-text indexes on mark text, owners, goods and services, and statements such as pseudo-marks, translations and disclaimers. It's built from the USPTO Open Data Portal's bulk files: the annual backfile (TRTYRAP) plus the daily application files (TRTDXFAP).
• Sync. sync/sync_daily.py runs at 6:30 a.m. and 6:30 p.m. It downloads any new daily files, loads them, and rebuilds the search indexes (about 12 minutes when there's something new). For each serial number, the record with the latest transaction date wins. It writes data/sync_status.json, and every report cites the date the register is current through.
• Search. search/clearance_search.py runs a battery of searches for a proposed mark (exact, all words, each word, spelling variants, compounds, pseudo-marks, near spellings) and ranks the candidates. search/domain_check.py checks domains with DNS and RDAP.
• The skill. skill/SKILL.md tells Claude how to run the searches, what to analyse, and how to write the report.

Requirements

• macOS (the scheduler uses launchd; on Linux, use cron and the same scripts)
• Python 3.9 or later; standard library only, nothing to install
• A free USPTO Open Data Portal API key, from data.uspto.gov
• About 100 GB of free disk space. The database is about 80 GB.
• Claude Code, or the Claude desktop app with this folder connected

Install

1. Put this folder in your home folder, not under Desktop, Documents or Downloads (macOS blocks scheduled jobs there):
   copy

   git clone https://github.com/jordanteague/tmrex.git ~/tmrex && cd ~/tmrex
   

2. Run setup. It asks for your API key and schedules the sync:
   copy

   bash setup.sh
   

3. Build the register. This takes several hours; if it's interrupted, run it again and it picks up where it stopped:
   copy

   caffeinate -i python3 -I sync/build_db.py
   

4. Install the skill:
   • Claude Code: mkdir -p ~/.claude/skills/tmrex && cp skill/SKILL.md ~/.claude/skills/tmrex/
   • Claude desktop app or claude.ai: upload skill/SKILL.md as a skill in Settings. In the desktop app, connect ~/tmrex as a folder in the session so the skill can reach the database.
5. Type /tmrex to see what to provide, or just ask: "Clear BLUE HERON for project management software."

If you already have a TrademarkBench register, setup clones it instead of building one (instant, using an APFS copy-on-write copy): it looks for ~/Desktop/TrademarkBench/local/uspto_full.db, or set BENCH_DB to its path.

Day to day

Task	Command
Clearance battery for a mark	python3 -I search/clearance_search.py clearance --mark "BLUE HERON" --classes 009,042
One custom register search	python3 -I search/clearance_search.py query '{"mark_query": "heron*", "classes": ["009"], "limit": 20}'
Full records	python3 -I search/clearance_search.py record 99123456
Domain status	python3 -I search/domain_check.py "blue heron" --names blueheron-app
See what a sync would load	python3 -I sync/sync_daily.py --dry-run
Run a sync now	python3 -I sync/sync_daily.py
Check how current the register is	cat data/sync_status.json
Read the sync log	tail -50 logs/sync.log
Stop the daily sync	launchctl bootout gui/$(id -u)/com.tmrex.uspto-sync

A failed file is retried on the next run. Each run looks back 7 days before the newest loaded file, so a file USPTO posts late is still picked up.

Limits

• The search matches words, spelling variants and near spellings. It can miss a word buried in a compound (GOWAVE for WAVE), sound-alikes that differ in their first letters, and foreign equivalents USPTO hasn't recorded; the skill covers these with follow-up searches and notes what's left.
• Words only: design marks and logos aren't compared.
• Prosecution history and design-search codes aren't stored.
• No state registrations or foreign registers.
• Reloading USPTO's next annual backfile once a year would correct any record a daily file missed.

Credits

search/search_register.py, search/normalize.py and sync/tm_parser.py come from TrademarkBench. The trademark data is public USPTO data and isn't covered by this repository's license.

License

MIT