Patent Checker

日本語版: README_ja.md

Patent Checker helps you explore published patents that may relate to your own software project, and to write down what you found as a dated, procedure-style report. It is meant to be run more than once on the same project — as an idea, during development, before a release, before each update: a later run searches only what is new, re-checks the patents you are monitoring, and tells you what changed since last time. It is built from two parts:

• an MCP server that fetches public patent data (EPO Open Patent Services and Google Patents) deterministically: searches, bibliographic records, claims, legal status, patent families, plus offline helpers, and
• an Agent Skill that drives your AI coding agent (Claude Code, Codex CLI, Cursor, Gemini CLI, GitHub Copilot CLI, OpenCode, OpenHands, Hermes Agent) through the judgment work: reading your codebase, translating its features into patent vocabulary, screening candidates and writing the report.

Patent Checker does not decide whether anything infringes a patent. Its reports contain observations, scope statements and open questions, never a verdict.

Status: stable release (v1.1).

Important notices

Please read these before you install anything. The tool asks you to acknowledge them once, and records that you did.

• What you learn here can be used against you. A report documents that you knew of specific patents on specific dates. In some jurisdictions, notably the United States, knowledge of a patent can support a claim of willful infringement and increased damages. Dated records of how you responded can also work in your favour. Patent Checker lets you choose where its records live and whether they are tracked in version control: make that choice deliberately.
• Search results are incomplete. No patent search is exhaustive. Coverage varies by country, language, publication stage and data source, and recent applications may be missing entirely. The absence of a document from a report is never evidence that no relevant patent exists.
• This is a search aid, not advice. Patent Checker does not replace a qualified patent attorney or agent, and its authors accept no responsibility for decisions made, or not made, on the basis of its output. The final judgment about your project, and any response to it, is yours.

There are two notices. The user notice (the three points above) is shown by the Skill to whoever runs an exploration, and by the installer up front; your agreement is recorded locally and checked before every run. The operator notice is for whoever starts the server: it covers data-source terms, credentials, and what the server stores, and the server refuses to start until it has been acknowledged.

How it works

your agent ──(Skill: judgment)──► patent-checker MCP server ──► EPO OPS
   │                                   │  deterministic fetch,      Google Patents
   │  reads your code, writes           │  normalisation, cache
   ▼  the report                        ▼
.patent-checker/reports/…        per-user document cache

1. The Skill reads your codebase, lists the technical features that could be claimed, translates them into patent vocabulary and builds search queries.
2. The MCP server runs the queries against EPO OPS, fetches the candidate documents (claims from Google Patents, bibliography and legal status from OPS) and caches every document so nothing is fetched twice.
3. The Skill screens the candidates in stages, maps claim elements to your features and writes a report: what was searched, what was found, how each candidate relates to your code, and what was not covered.
4. The Skill also keeps a ledger of the exploration in your project (features, queries, screened families, monitored patents and a snapshot of their status). The next run starts from it: the server compares the stored snapshots with the current records, and the new report opens with what changed.

The server only ever receives public patent data: search expressions, publication numbers, family identifiers, dates, and status snapshots it returned earlier. Your source code, your project description and the ledger never leave your machine: the analysis happens inside your agent.

Prerequisites

• An AI coding agent that supports Agent Skills and MCP: Claude Code, OpenAI Codex CLI, Cursor, Gemini CLI, GitHub Copilot CLI, OpenCode, OpenHands or Hermes Agent.
• Python 3.12 or newer and uv for the command-line tool and installer (pipx or a plain pip install in a virtual environment work too). Python is also the easiest way to generate the bearer token below.
• Docker with Compose v2, if you run the server in a container (recommended). Without Docker the server runs as an ordinary process.
• An EPO Open Patent Services account (recommended). Register at https://developers.epo.org/, then create an app to obtain a consumer key and secret. Apply early: approval is manual and typically takes about a business day. Without it the server runs in degraded mode.

The bearer token

The server listens on HTTP, and every MCP client must present a bearer token on every request: a secret string that only your server and your own agents know. It is what stops any other program on the machine, or on the network if you ever expose the port, from using your server and your EPO OPS quota. There is no account or sign-up behind it; you make the token yourself, once, and give the same value to the server and to the installer.

Generate one (any 32 random bytes will do):

python3 -c "import secrets; print(secrets.token_urlsafe(32))"      # macOS, Linux

python -c "import secrets; print(secrets.token_urlsafe(32))"       # Windows

Where it goes:

• Docker Compose: into the file secrets/server_token.txt next to compose.yaml.
• Without Docker: into .env as PATENT_CHECKER_SERVER_TOKEN=… (or a file named by PATENT_CHECKER_SERVER_TOKEN_FILE).
• Your agents: patent-checker install asks for it (or reads it from --token-file), and writes it into each agent's MCP configuration.

Without a token the HTTP server does not start. The one exception is patent-checker serve --transport stdio, which has no token because the agent starts the server as a child process running with your own permissions; that is a special case for a single local client, not the default. Treat the token like a password: keep the files that hold it private, and rotate it by writing a new value and re-running patent-checker install.

Without an EPO OPS account (degraded mode)

If you leave the OPS credentials unset (empty secrets/ops_key.txt and ops_secret.txt, or no PATENT_CHECKER_OPS_KEY in .env), the server starts in degraded mode and says so at start-up and in server_status.

• Still works: fetching the claims of a document from Google Patents when your agent already knows its publication number, publication-number normalisation, the offline helpers (deduplication, batch verification).
• Unavailable: patent searches, bibliographic records, legal status and patent families. The Skill switches to its degraded procedure, in which your agent finds candidate publication numbers through its own web search; coverage drops, and the report says so.

Degraded mode is meant for a first look. For a real exploration, get an OPS account and restart the server with the credentials in place.

Install the server

Both routes need the operator notice acknowledged and a bearer token.

Option A: Docker Compose (recommended)

mkdir patent-checker-server && cd patent-checker-server
curl -fsSLO https://raw.githubusercontent.com/xhighhongo41/patent-checker/main/compose.yaml
mkdir secrets

# 1. EPO OPS credentials (or leave both files empty for degraded mode)
printf '%s' 'YOUR_OPS_CONSUMER_KEY'    > secrets/ops_key.txt
printf '%s' 'YOUR_OPS_CONSUMER_SECRET' > secrets/ops_secret.txt

# 2. The bearer token (see above)
python3 -c "import secrets; print(secrets.token_urlsafe(32))" > secrets/server_token.txt
chmod 600 secrets/*.txt

# 3. Read the operator notice (add --lang ja for Japanese), then acknowledge it
docker compose run --rm patent-checker serve --show-operator-notice
echo 'PATENT_CHECKER_OPERATOR_CONSENT=1.0' >> .env

# 4. Start
docker compose up -d
curl -fsS http://127.0.0.1:8642/health

On Windows, create the three files with PowerShell instead of printf (Set-Content writes a byte-order mark that would become part of the token, so use .NET directly):

New-Item -ItemType Directory -Force secrets | Out-Null
New-Item -ItemType File -Force secrets\ops_key.txt, secrets\ops_secret.txt | Out-Null   # empty: degraded mode
$token = python -c "import secrets; print(secrets.token_urlsafe(32))"
[IO.File]::WriteAllText("$PWD\secrets\server_token.txt", $token)

The image is published as ghcr.io/xhighhongo41/patent-checker and docker.io/xhighhongo41/patent-checker with tags X.Y.Z, X.Y and latest. The container binds 0.0.0.0 internally, but Compose publishes the port on your machine's loopback only (127.0.0.1:8642); the server accepts the localhost and 127.0.0.1 Host headers and nothing else. Fetched documents, search results and the request log live in the named volume patent-checker-data. The container runs as an unprivileged user on a read-only filesystem with all Linux capabilities dropped; if your Docker engine rejects one of those settings, the four lines under # Hardening in compose.yaml can be removed without changing what the server does.

The secret files are read once at start; empty OPS files mean "not configured" (degraded mode). If you run Docker Engine on Linux under an account that is not UID 1000, see the note on file ownership in secrets/README.md.

To serve a LAN or a VPN (several machines, one server), see docs/deploy-lan.md: it adds a TLS-terminating proxy in front of the same container.

Option B: without Docker

uv tool install patent-checker          # or: pipx install patent-checker
mkdir patent-checker-server && cd patent-checker-server
curl -fsSLO https://raw.githubusercontent.com/xhighhongo41/patent-checker/main/.env.example
cp .env.example .env

Then edit .env:

1. Put the EPO OPS key and secret in PATENT_CHECKER_OPS_KEY and PATENT_CHECKER_OPS_SECRET (leave them empty for degraded mode).
2. Generate a bearer token (see above) and put it in PATENT_CHECKER_SERVER_TOKEN.
3. Read the operator notice with patent-checker serve --show-operator-notice (add --lang ja for Japanese) and set PATENT_CHECKER_OPERATOR_CONSENT=1.0.

A minimal .env looks like this:

PATENT_CHECKER_OPS_KEY=YOUR_OPS_CONSUMER_KEY
PATENT_CHECKER_OPS_SECRET=YOUR_OPS_CONSUMER_SECRET
PATENT_CHECKER_SERVER_TOKEN=the-token-you-generated
PATENT_CHECKER_OPERATOR_CONSENT=1.0

Start the server with patent-checker serve. It speaks Streamable HTTP on http://127.0.0.1:8642/mcp. The .env file is looked up in the directory you start from and its parents, stopping short of your home directory and the filesystem root; variables already in the environment win. Data goes to the per-user data directory (~/.local/share/patent-checker on Linux and macOS, %LOCALAPPDATA%\patent-checker on Windows). If the port is already taken, the server stops with a message; pick another with --port and register that URL with your agents.

--transport stdio runs the server without a network port and without a token, for a single local client only; it then runs with your user's permissions, which is why HTTP is the default. How to register it with an agent is in docs/mcp-clients.md.

Install the Skill and connect your agent

One command installs the command-line tool, shows the user notice, copies the Skill into the places your agents read, and registers the MCP server with them:

curl -LsSf https://raw.githubusercontent.com/xhighhongo41/patent-checker/main/install.sh | sh

powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/xhighhongo41/patent-checker/main/install.ps1 | iex"

Prefer to read the script first? Fetch it with | more instead of | sh, or skip it entirely: the script only installs uv and the tool, and then tells you to run the installer yourself.

uv tool install patent-checker
patent-checker install

patent-checker install:

1. shows the user notice and asks for your agreement (recorded in ~/.config/patent-checker/consent.json with the date and notice version; you are asked again only when the notice changes);
2. asks for the bearer token (or reads it from --token-file);
3. copies the Skill into ~/.agents/skills/patent-checker/ (read by Codex CLI, OpenCode, Cursor, Gemini CLI and Copilot CLI) and into the private directories of the agents that need one (~/.claude/skills/, ~/.openhands/skills/, ~/.hermes/skills/);
4. registers the server (http://127.0.0.1:8642/mcp unless you pass --url) with each detected agent, using the agent's own mcp add command where one exists (Claude Code, Gemini CLI, Copilot CLI, Codex CLI with --token-env), editing the agent's JSON or TOML configuration otherwise (Cursor, OpenCode, Codex CLI), and printing a snippet to paste for the agents whose configuration it will not touch (OpenHands, Hermes Agent);
5. prints a table of what it did per agent. The exit code is non-zero if any step failed; a printed snippet is not a failure.

Useful options: --list-agents (what would be installed where), --dry-run (do everything except write), --agent claude-code --agent cursor (instead of auto-detection; --agent all for every supported agent), --scope project (install into the current project instead of your home directory), --lang ja (Japanese notice), --no-skill / --no-mcp. Re-running the installer is safe: it refreshes the Skill copies and updates the server entry in place, and it never overwrites the backup (.bak) it made of your original configuration file on the first run. The one exception is Codex CLI, whose existing entry is left alone; edit ~/.codex/config.toml to change it.

Passing the bearer token without exposing it

The installer never takes the token as a command-line argument, so it does not land in your shell history. Give it one of:

patent-checker install --token-file /path/to/secrets/server_token.txt
# or
export PATENT_CHECKER_SERVER_TOKEN="$(cat /path/to/secrets/server_token.txt)"
patent-checker install
# or just run it: the installer prompts for the token without echoing it

By default the token value is written into each agent's own configuration file, which is then made readable by you alone (on Windows the file keeps its usual permissions). With --token-env the installer does not need the value at all: it writes a reference to the PATENT_CHECKER_SERVER_TOKEN environment variable, in the notation each agent expands, and you export that variable before starting the agent. Copilot CLI and OpenHands do not document such references, so --token-env prints a snippet for them instead. When an agent's own CLI is used for registration, the token appears in that process's arguments for the duration of the call.

With --scope project, the configuration file that holds the token is written inside your project directory. The installer reminds you: do not commit it.

Registration commands and snippets that the installer prints never contain the token; they use ${PATENT_CHECKER_SERVER_TOKEN} or a placeholder.

If the server answers 401

A 401 means the token the agent sends is not the one the server was started with. Compare the two: the server's is in secrets/server_token.txt or .env; the agent's is in the configuration file from the table above (Claude Code shows it with claude mcp get patent-checker). Make sure a placeholder such as <token> did not get registered verbatim. A server that is not running produces a connection refused error, not a 401, so the two are easy to tell apart.

Registering by hand

The exact command or file for each agent, the GitHub Copilot coding agent, and the stdio form are in docs/mcp-clients.md.

Phone-only and cloud agents

Commit the Skill into your repository as .agents/skills/patent-checker/ (patent-checker install --scope project --no-mcp, then add the directory to git), and run the server somewhere the cloud agent can reach over HTTPS (docs/deploy-lan.md). The consent record is per user; a cloud agent whose home directory is reset records it again on the next run, or use patent-checker consent record --project to keep it with the project.

Usage

Ask your agent for a prior-art exploration of the project it is working in:

Use the patent-checker skill to explore prior patents related to this project.

The Skill first checks the consent record, asks (once per project) whether .patent-checker/ should be added to .gitignore, then works through the steps above and writes the report to .patent-checker/reports/report-<target>-<YYYYMMDD-HHMM>.md. Reports are never overwritten; a later run of the same project produces a new, dated file. Keep the server running for the whole session.

Scale and cost. One exploration of a medium-sized project involves reading your code, one or more search rounds, and a staged screening of a few dozen candidate documents. Expect it to consume on the order of a million tokens of agent traffic (measured with Claude Code on a medium-sized project during development), most of it in screening; the Skill hands the first screening stage to smaller models where your agent supports delegation. One run is not exhaustive: repeated runs, different query vocabularies and a professional search will each find things a single run does not.

Running it again

Ask again whenever the project has moved on, or when the date of the next check in the report has come:

Use the patent-checker skill to follow up on the earlier exploration of this project.

The Skill finds the ledger under .patent-checker/ledger/<target>/ and proposes one of two kinds of run:

• A follow-up run reads what changed in your code since the explored commit, runs the earlier queries only for patents published since then (new queries cover new features), screens only families it has not seen before, re-checks every monitored patent, and re-reads a document only if your feature, its claims, or an earlier reading changed. The report describes the current state of everything and opens with "Changes since the previous exploration": changes of your project, new documents, changes of legal status and of claims, observations that changed and why, and the documents that were checked and found unchanged.
• A monitoring run only re-checks the monitored patents and writes a short update-<target>-<YYYYMMDD-HHMM>.md. It costs a small fraction of an exploration, and says plainly that nothing was searched.

Nothing is downloaded twice: claims stay in the cache for good, legal status is refreshed after a week and families after a month, and every result says when it was fetched, so a report can state what its facts are "as of". Measured during development, three weeks after the first exploration of a medium-sized project: the 17 stored queries returned 11 hits for the new publication window instead of 712 for all time, 8 of them in families that had been screened already, and all 25 monitored patents were re-checked with about fifty upstream requests; a monitoring run made the same day needed none. If you explored a project with an earlier version, the Skill offers to import the latest report into a ledger first.

Configuration

Environment variables read by the server (patent-checker serve) and, where noted, by the CLI. See .env.example for the same list with comments. The CLI's --host and --port options override PATENT_CHECKER_SERVER_HOST and PATENT_CHECKER_SERVER_PORT.

.env is read by every patent-checker command from the directory it runs in or a parent of it, never from your home directory or the filesystem root. With Compose, the credentials come from the files under secrets/, and the .env next to compose.yaml provides PATENT_CHECKER_OPERATOR_CONSENT (plus, optionally, PATENT_CHECKER_IMAGE, PATENT_CHECKER_PORT and PATENT_CHECKER_LOG_LEVEL). Other variables can be added under environment: in compose.yaml.

Cache and clean-up

Fetched patent documents are kept once, in the per-user shared cache, with a per-kind expiry; search results and the request log stay with whoever ran them (the server's data directory, or the project's .patent-checker/ when the CLI is used from the project).

• patent-checker cache status shows what is cached, where, and how much of it has expired or is broken.
• patent-checker cache clear [--kind KIND] [--older-than DAYS] [--pub …] [--expired] [--broken] lists what it would delete; add --yes to delete. KIND is one of biblio, claims, legal, family, gp, search, searchbib.
• patent-checker clean lists the project's traces (search cache, request log, leftovers of older layouts) and, with --yes, deletes them. Reports, your agent's exploration notes, the ledger and the consent record are kept unless you add --include-artifacts or --include-consent; --shared extends the clean-up to the shared cache and the server's data directory.
• patent-checker ledger status summarises the ledgers of the project (runs, monitored patents, checks that are due) and patent-checker ledger check verifies one; both only read. The ledger is written by your agent, lives with the reports, and is covered by the same choice of tracking it in git or not: it is a dated record of what you knew, just as a report is.
• A publication number without its kind code (EP1234567) means "whatever is current", so claims fetched that way expire like a family instead of being kept for good; with the kind code (EP1234567B1) they never expire.
• With Compose, docker compose down -v removes the server's volume, cache and all.

Nothing is deleted without --yes.

Security model

• The server binds the loopback interface by default and requires a bearer token on every request, compared in constant time. Host and Origin headers are validated against the names you configured (no wildcards), which is what stops a DNS-rebinding attempt from reaching a loopback server; a page served from loopback itself is not blocked by that check, and is kept out by the token.
• It only ever connects to ops.epo.org and patents.google.com, from the server and from the command-line tool alike, and only writes below its own data directory. It has no code-execution tools.
• It receives public patent data only — search expressions, publication numbers, family identifiers, dates, and status snapshots it returned earlier — and refuses oversized requests. Your code, your documents and the ledger are handled by your agent, on your machine. Authenticated clients can see where the server keeps its data (server_status); nothing else about the host is exposed.
• Incoming requests are rate-limited, and upstream requests are paced, serialised one at a time and cached so that a runaway agent cannot hammer the data sources.
• GET /health is unauthenticated for container health checks and returns only a status and the version.
• Credentials are read from the environment or from files, are never logged, and never appear in the startup banner or in error messages.
• The container runs as an unprivileged user on a read-only filesystem with all capabilities dropped; only /data is writable.
• --transport stdio has no authentication: use it only for a single local client, aware that the server then runs with your permissions.
• Serving a LAN or the internet is possible but is your responsibility: the transport is plain HTTP, so put a TLS-terminating reverse proxy in front (docs/deploy-lan.md), set PATENT_CHECKER_SERVER_ALLOWED_HOSTS, and treat the token as the only thing between the network and your OPS quota.
• The Skill treats everything the server returns, patent text included, as data to analyse, not as instructions to follow.
• Releases are built and published by CI from a version tag; dependencies and the container image are scanned for known vulnerabilities on every change. To report a security problem, see SECURITY.md.

Data sources and fair use

• EPO Open Patent Services is used under the terms attached to your registered application, including its weekly fair-use quota. The server spaces its requests and honours OPS throttling headers, but it cannot know how many other clients share your credentials. Legal-status information comes from OPS and is the authoritative value in reports. That pacing works inside one process: the server is one, but every patent-checker command you run from a shell is its own. When you drive the CLI by hand or from a script, put several queries into one plan-check call and leave some fifteen seconds between search commands; an error mentioning 403 means OPS throttled the search service, and waiting a quarter of an hour clears it.
• Google Patents is fetched one document page at a time, at a pace and volume comparable to a person reading in a browser, and every page is cached so it is not fetched twice. The search endpoint is deliberately not used: Google's robots.txt permitting a path is not the same as its terms of service permitting bulk collection.
• Patent documents are copyrighted by their applicants. Reports quote only the claim passages they discuss, with the source, and never redistribute full texts; the cache stays on your machine and is not part of any report.

Notes

• US publication numbers changed length in 2026. EPO's DOCDB representation of US pre-grant publications is 10 digits up to 2025 (US2007016547A1) and 11 digits from 2026 (US20260024003A1). Google Patents spells all years with 11 digits. The tools normalise both spellings; when you type a number by hand, either form is accepted.
• A document whose legal status comes back with no events at all is reported as such (events: [] with a note), which is not the same as "not found".
• When a notice changes, the Skill asks for consent again, or the server asks for a fresh acknowledgement; until then nothing is asked twice.

Updating

Docker Compose: docker compose pull && docker compose up -d.

Without Docker: stop the running server, then

uv tool upgrade patent-checker      # or: pipx upgrade patent-checker
patent-checker serve

Tool and Skill on your machine: uv tool upgrade patent-checker && patent-checker install. Re-running the installer refreshes the Skill copies and the server entry in each agent's configuration. Keep the server and the Skill at the same version: both are updated by the same command, and a Skill from a newer release may call tools an older server does not have.

Uninstalling

• Agents: remove the patent-checker entry from each agent's MCP configuration (claude mcp remove patent-checker, gemini mcp remove patent-checker, or delete the entry from the file in the table above), and delete the Skill directories ~/.agents/skills/patent-checker/, ~/.claude/skills/patent-checker/, ~/.openhands/skills/patent-checker/, ~/.hermes/skills/patent-checker/ (and their .agents/skills/ and .claude/skills/ counterparts in projects where you used --scope project).
• Server: docker compose down -v (container, cache and all) or uv tool uninstall patent-checker (plus the per-user data directory, ~/.local/share/patent-checker or %LOCALAPPDATA%\patent-checker, if you want the cache gone too).
• Records: patent-checker clean --yes --include-artifacts --include-consent in a project removes its reports, notes and consent record (run it before uninstalling the tool); ~/.config/patent-checker/consent.json holds the per-user consent record.

Changelog

• v1.1 (2026-09): built for repeated runs. A per-project ledger carries an exploration from one run to the next; follow-up runs search only newly published patents and screen only new families; the new watch_check tool compares the legal status and family of monitored patents with the previous run; reports open with "Changes since the previous exploration", and a short monitoring update exists for status-only checks. Every result says when it was fetched (fetched_at), claims results name the publication actually read (pub_docdb), claims requested without a kind code no longer stay cached for good, dedup_families marks known families, usage_report takes a start time, new patent-checker watch and patent-checker ledger status|check commands. The container image applies the distribution's security updates when it is built.
• v1.0 (2026-09): stable release. Constant-time bearer-token check, .env never read from your home directory or the filesystem root, request size limits, hardened container (read-only, no capabilities), LAN/TLS example with Caddy, installer improvements (--token-env needs no token value, owner-only configuration files, backups never overwritten, non-zero exit on failure, Codex CLI registration), cache and parser fixes (concurrent writes, CPC symbols, empty legal status, non-UTF-8 pages), upstream calls retried once on 401/429/503 and never blocking unrelated tools, every command reads .env, plain error messages instead of tracebacks, Python 3.13, security scanning in CI, listing in the MCP Registry, reorganised documentation.
• v0.5 (2026-09-06): first public release. Docker image (GHCR, Docker Hub) and Compose deployment with file-based secrets, patent-checker install for eight agents, bootstrap scripts, PyPI package, /health, log level setting.
• v0.4 (2026-09-05): per-user shared document cache with per-kind expiry and self-repair, cache status / cache clear / clean, publication number normalisation for the 2026 change in US numbers.
• v0.3 (2026-09-04): MCP server (FastMCP 4, Streamable HTTP, bearer token, Host/Origin validation, outbound allowlist, rate limits), operator notice, first file cache.
• v0.1–v0.2 (2026-08): development versions; core library, CLI and the Agent Skill with its consent gate and report template.

License

Apache-2.0. See LICENSE.

Questions and bug reports: GitHub Issues.