Skip to main content
Glama
patent_research

patent_research

Official

Patent Research Hub

Public preview 0.1.0 — Apache-2.0. Published by Brightline IP.

Patent Research Hub retrieves public US patent and legal sources for an AI assistant or a script, keeps a local, provenance-tracked record of what it retrieved, and exports source-linked evidence packets. It is a deterministic retrieval layer: it makes no model calls, accepts no arbitrary SQL or URLs, and never presents a search page as a complete population or a legal conclusion.

It is meant for patent practitioners and researchers who want an MCP client (Claude Desktop, Claude Code, Codex and others) or a command line to pull Federal Circuit opinions, PTAB decisions, USPTO file-wrapper records, EPO bibliography, MPEP/35 USC/37 CFR references and Google-hosted patent statistics with their provenance intact. Read DISCLAIMER.md: output is not legal advice and must be checked against the official record.

Direct transport has not been verified against live providers. In this release the default direct transport (the Hub calling USPTO, EPO, GovInfo, CourtListener and BigQuery itself) is covered only by offline tests with mocked provider responses. Request shapes follow each provider's documentation and the n8n workflows the Hub was originally built and live-tested with, but expect to find provider-side differences. Please report them (see SECURITY.md for security issues).

What it does

Area

Sources

Key needed

Court and Board decisions

Federal Circuit RSS (opinions/orders), PTAB precedential/informative index, PTAB statistics, Director Review pages

No

Official notices

USPTO Official Gazette, e-FOIA reading rooms, examination guidance index

No

USPTO Open Data Portal

Patent file-wrapper search, PTAB trial and ex parte appeal decisions, application document lists, identity-checked original PDFs

USPTO_API_KEY

EPO Open Patent Services

Bibliographic lookup by EPODOC publication number

EPO_OPS_CONSUMER_KEY + EPO_OPS_CONSUMER_SECRET

GovInfo

USCOURTS opinion search

GOVINFO_API_KEY

CourtListener

Opinion search (disabled by default; requires accepting Free Law Project terms)

COURTLISTENER_TOKEN + config

Google BigQuery

Fixed CPC publication-year trend and hosted-PatEx examiner/art-unit outcome aggregates; dry-run byte estimates by default

GOOGLE_APPLICATION_CREDENTIALS + BIGQUERY_PROJECT_ID

Reference library

MPEP chapters, 35 USC and 37 CFR consolidations, FRCP: a manifest of official URLs and SHA-256 pins plus a fetch command

No

Local datasets

PatentsView annualized grant cohort (you download and prepare it), PatEx (via BigQuery)

No

Link-only

WIPO PATENTSCOPE and IP Statistics, Google Patents, FreePatentsOnline, Patent Bots, Juristat, Patlytics

No automated retrieval

CPC scheme lookups (src/epo-cpc.mjs) and GovInfo opinion PDF download (src/govinfo-pdf.mjs) exist as module functions that other tools can import. No CLI command or MCP tool calls them, so they are not part of what the command line or an assistant can do today.

Every retrieval is saved as an immutable response snapshot plus normalized records in a local SQLite database, so earlier results stay searchable offline. Reports are exported as JSON, Markdown and HTML (with CSV for statistics, and a chart only when Python with matplotlib is available; without them the report is still written, with a warning and no chart). Report tools put warnings first in their result; if writing the files fails after a paid BigQuery query, the result is still returned with an exportError field.

Related MCP server: Claude Patent Creator MCP Server

Requirements

  • Node.js 24.13 or newer (the Hub uses the built-in node:sqlite module). Windows, macOS or Linux; the developers use Windows, and other systems have not been tested.

  • Optional: Python 3.12 or newer with requirements.txt (matplotlib). Python is needed only for the charts in reports and for PatentsView preparation. Reports still work without it: they are written with a warning and no chart.

  • Optional: your own provider credentials (table above). Nothing is needed for the keyless sources.

Install on Windows (easiest)

  1. Download patent-research-hub-0.1.0.zip from the release page and check its SHA-256 against the value in the release notes. In PowerShell: Get-FileHash .\patent-research-hub-0.1.0.zip -Algorithm SHA256

  2. Right-click the zip, choose Extract All, open the extracted folder and double-click Setup.cmd. The script is not code-signed, so Windows SmartScreen may warn you; choose More info, Run anyway only if the checksum matched.

  3. Setup checks for Node.js (required) and Python (optional: charts in reports, which are still produced without it, and PatentsView preparation). If one is missing it explains why, shows the exact winget command and installs it only if you answer yes. Both are free. No administrator rights, Docker or API key are needed.

  4. Setup copies the program to %LOCALAPPDATA%\Brightline\PatentResearchHub, runs npm ci --omit=dev from package-lock.json, and creates a separate data folder, %LOCALAPPDATA%\Brightline\PatentResearchHub-Data, that only your Windows account can read. It holds your database, exports, config.json (BigQuery stays in dry-run mode) and an empty keys.env. Setup never asks for a key and changes no AI assistant's configuration.

  5. When it finishes, Setup runs the check and opens a getting-started guide. Sources without a key show "not configured"; the keyless sources work straight away.

Start Menu, folder Patent Research Hub:

  • Check: which sources are ready (setting names only, never key values; no network call).

  • Data folder: opens the data folder, where keys.env and config.json are.

  • Documentation: the getting-started guide (where keys go, how to connect an assistant).

  • Repair: reinstalls the npm packages, may offer any missing prerequisite (Node.js, optional Python) with your yes, re-applies the protection on keys.env, refreshes the optional chart packages if the private .venv exists, and regenerates Hub.cmd, the MCP setup files, the getting-started guide and the shortcuts. It does not replace program files (run Setup.cmd from a newer download for that) and does not touch your data, settings or key values. It does not open the guide in Notepad.

  • Uninstall: removes the program and shortcuts; keeps the data folder unless you choose to delete it. It ends by saying whether the program folder and the data folder were deleted or kept.

Keys. Open keys.env from the data folder in Notepad, paste each value after its = sign, save, and run Check. Registering for the USPTO, EPO OPS, GovInfo and CourtListener keys is free (each provider sets its own terms and limits); BigQuery can cost money on a billing-enabled Google Cloud project, which is why the Hub only makes free dry-run estimates until you change config.json. Nothing leaves your computer until you use a source; then your query terms, document numbers and that provider's key go to that provider only.

AI assistants. Setup writes MCP-setup.json (Claude Desktop, Claude Code) and MCP-setup.codex.toml (Codex) in the program folder. Merge the patent_research entry into your client's configuration yourself. The entry starts Node with --env-file-if-exists=<data folder>\keys.env, so keys stay in that one file and out of client settings. A variable that is already set in your Windows environment takes precedence: Node's --env-file does not override it, so a stale USPTO_API_KEY set elsewhere will win over the one in keys.env. From a command prompt, "%LOCALAPPDATA%\Brightline\PatentResearchHub\Hub.cmd" doctor runs the CLI with the same folders and keys.

Running Setup again repairs or updates the installation. Setup.cmd -DryRun shows what it would do without changing anything; -InstallRoot, -DataRoot and -ShortcutFolder choose other folders. If Hub processes from this installation are running (an AI assistant usually starts them), Setup explains and asks before stopping them (default: no). An update replaces the program folder with the new download: anything you saved inside the program folder is deleted. Keep your own files in the data folder or elsewhere. The program folder and the folder you extracted the download into must not contain one another; Setup refuses that before copying anything. If the package download fails during an update, the Hub is unusable until Setup is run again (Setup.cmd, or the Repair shortcut).

Install from source

git clone <this repository> patent-research-hub
cd patent-research-hub
npm ci
# optional, for charts:
python -m venv .venv && .venv/Scripts/python -m pip install -r requirements.txt   # Windows
# python3 -m venv .venv && .venv/bin/pip install -r requirements.txt               # macOS/Linux

npm ci downloads the four npm dependencies and their transitive packages. No service, scheduled task or background process is installed.

Quick start

node src/cli.mjs doctor                         # which sources are configured (names only, never values)
node src/cli.mjs refresh --source cafc          # recent Federal Circuit opinions and orders (no key)
node src/cli.mjs search --query "obviousness"   # search what you have saved locally
node src/cli.mjs query --source odp --query 'applicationMetaData.inventionTitle:"robot"' --limit 5   # needs USPTO_API_KEY
node src/cli.mjs bigquery-estimate --cpc G06N --from 2024 --to 2024                                  # dry run only
node src/cli.mjs references-fetch --source usc35                                                     # official PDF, hash-checked

To load keys from a local file: copy .env.example to .env, fill in what you have, and run node --env-file=.env src/cli.mjs doctor. --env-file does not override a variable that is already set in your environment (the existing value wins). node src/cli.mjs with no arguments lists every command.

A source without its key returns {"status":"not_configured", "missing":[...], "register":"<signup link>"} (CLI exit code 2) instead of failing.

MCP client registration

The MCP server speaks stdio and exposes 36 tools (research_doctor, research_sources, research_query, research_refresh, research_search, research_record, BigQuery, examiner, reference-library, report, checkpoint and backup tools). Example for Claude Desktop or any client that uses the mcpServers format — use absolute paths:

{
  "mcpServers": {
    "patent_research": {
      "command": "node",
      "args": ["/absolute/path/to/patent-research-hub/src/mcp.mjs"],
      "env": {
        "USPTO_API_KEY": "your-key",
        "PATENT_RESEARCH_DATA_DIR": "/absolute/path/to/patent-research-data"
      }
    }
  }
}

Allow a generous tool timeout (about 4 minutes) for BigQuery and original-document tools. The optional skill text in skills/patent-research/SKILL.md tells an agent how to use the tools conservatively.

Configuration

Settings come from an optional config.json in the install folder (copy config.example.json; it is git-ignored) or the file named by PATENT_RESEARCH_CONFIG. Credentials are only read from environment variables.

Setting / variable

Default

Meaning

transport / PATENT_RESEARCH_TRANSPORT

direct

direct: the Hub calls providers itself. n8n: optional loopback n8n holds the credentials (extras/n8n).

USPTO_API_KEY

—

USPTO Open Data Portal key. Register.

EPO_OPS_CONSUMER_KEY, EPO_OPS_CONSUMER_SECRET

—

EPO OPS application credentials; the Hub obtains and caches OAuth tokens in memory. Register. EPO_ACCESS_TOKEN is accepted as a short-lived alternative.

GOVINFO_API_KEY

—

GovInfo (api.data.gov) key. Register.

COURTLISTENER_TOKEN

—

CourtListener API token. API docs.

courtlistenerEnabled, courtlistenerInternalUseAllowed

false, false

Both must be true before CourtListener is used. Set the second only after reading the Free Law Project allowed-use terms.

GOOGLE_APPLICATION_CREDENTIALS

—

Path to a Google service-account key or gcloud auth application-default login credential file.

BIGQUERY_PROJECT_ID / googleBigQueryProjectId

—

Your own Google Cloud project (queries are run and billed there).

googleBigQueryDryRunEnabled

true

Free dry-run byte estimates.

googleBigQueryExecutionEnabled

false

Allows capped data queries. Also needs googleBigQueryBillingMode.

googleBigQueryBillingMode

disabled

sandbox-no-billing (BigQuery sandbox, no billing account) or billing-enabled-capped (you accept possible charges up to the caps).

googleBigQueryPerQueryMaxBytes

26843545600 (25 GiB)

Sent as maximumBytesBilled; a dry-run estimate above it is refused before execution.

googleBigQueryMonthlyTrackedMaxBytes

274877906944 (256 GiB)

Calendar-month budget tracked locally for queries made by this tool only.

maxRequestsPerHour

120 per keyed provider (USPTO, EPO, GovInfo, CourtListener, BigQuery); 0 = unlimited

Local request-count limit per provider and process. Not the provider's own quota.

directMinIntervalMs

USPTO/EPO/CourtListener/BigQuery/public 1000 ms, GovInfo 3600 ms

Minimum spacing between requests to one provider.

requestTimeoutMs

25000

Per-request timeout (PTAB originals 60 s, application originals and BigQuery 180 s).

ptabOriginalsEnabled, patentDocumentsEnabled, patentDocumentOriginalsEnabled

true

Allow the USPTO document list and identity-checked original downloads.

PATENT_RESEARCH_DATA_DIR / dataDir

<install>/data

Local database, backups, checkpoints, staged originals, reference PDFs.

PATENT_RESEARCH_EXPORTS_DIR / exportsDir

<install>/exports

Exported reports.

RESEARCH_PYTHON / pythonExecutable

.venv, then Python on PATH

Charts in reports (optional; without Python the report has no chart and says so) and PatentsView preparation.

RESEARCH_GET_FILES_CLIENT / getFilesClientModule

—

Optional Get Files integration for archiving and full-text reading.

RESEARCH_CORPUS_CLIENT / corpusClientModule

—

Optional corpus client for PatentsView intake.

corpusIntakeRoot

<dataDir>/intake

Where verified originals are staged.

Network behaviour on the direct transport: HTTPS only, fixed provider hosts and paths (the same validation the n8n workflows use), no redirects followed except the one identity-checked USPTO signed download, response size caps, timeouts, and retries with backoff (Retry-After honoured, at most 5 s) only for idempotent GET requests. POST searches and BigQuery calls are never retried.

BigQuery cost warning

BigQuery charges by bytes scanned. A single CPC trend query over the Google Patents publications table can scan tens of GiB (one 2021–2025 G06N query scanned about 16.5 GiB). At on-demand prices that is a fraction of a dollar per query, and Google currently gives every project a free monthly allowance, but prices, free tiers and your project's billing state are your responsibility. The Hub therefore:

  • runs only fixed, parameterized SQL; it never accepts SQL from a caller;

  • defaults to dry runs, which return a byte estimate and process no data;

  • runs a data query only when googleBigQueryExecutionEnabled is true and a billing mode is acknowledged, after a fresh dry run under googleBigQueryPerQueryMaxBytes, with that ceiling sent to Google as maximumBytesBilled, and only while the local monthly ledger stays under googleBigQueryMonthlyTrackedMaxBytes;

  • cannot see queries you run elsewhere (console, other tools), so the monthly ledger is a floor, not your bill. Google enforces only the per-query maximumBytesBilled; the monthly cap is a local ledger, and reconciling an interrupted reservation is an operator-only CLI step (bigquery-reconcile, never an MCP tool).

Use a BigQuery sandbox project (no billing account) if you want a hard stop. See docs/BIGQUERY_SETUP.md and Google's pricing page.

Provider terms

Each user registers for and uses their own credentials and must follow each provider's terms and rate limits. The Hub spaces requests (directMinIntervalMs) and stops at maxRequestsPerHour per provider (default 120, this process only); both are local safeguards, not your provider's quota (see docs/INTEGRATION.md): USPTO Open Data Portal, EPO OPS terms and fair-use charter, GovInfo API, CourtListener allowed API usage (membership-based; organizational or commercial use may need Free Law Project's agreement), Google BigQuery pricing. WIPO PATENTSCOPE prohibits robots and bulk downloads, and Patent Bots prohibits systematic downloading; the Hub only generates links for those services. More in docs/FREE_SOURCES.md and docs/OPTIONAL_CONNECTORS.md.

Data locations

Everything stays on your computer unless you call a provider:

  • <dataDir>/research.sqlite — saved records, response snapshots, run history, BigQuery ledger, aggregate cache.

  • <dataDir>/backups, <dataDir>/recovery — backup and restore-stage output.

  • <dataDir>/checkpoints — research handoffs (research_checkpoint).

  • <dataDir>/intake — verified original PDFs staged for an optional corpus.

  • <dataDir>/references — official reference PDFs fetched by references-fetch.

  • <dataDir>/patentsview — PatentsView snapshots you prepare.

  • <exportsDir>/<uuid>/ — exported reports.

There is no automatic retention or pruning. Response snapshots, backups, checkpoints, staged originals and exported reports accumulate until you delete them yourself (the backup command never removes earlier backups). Check the data folder's size now and then, and delete old backup, export or checkpoint files yourself. The Hub has no command that prunes records or snapshots inside research.sqlite.

No provider data, reference PDFs or credentials ship with the software.

Security model

  • Credentials are read from environment variables at call time, sent only to their own provider host, held in memory only (OAuth tokens included), and never written to the database, reports, checkpoints or logs. doctor reports which variables are set, never their values. Error messages are redacted.

  • The Hub contacts only fixed HTTPS provider endpoints; caller input cannot choose a host, path or SQL text.

  • Retrieved text is untrusted data. Tool results label coverage limits; a search page is never a population.

  • The optional n8n transport requires a bearer file you create and talks to n8n over loopback http by default. The one exception is an explicit opt-in: set n8nRemoteHost to the exact hostname of an n8n you operate and n8nBaseUrl to https://<that host> (a bare origin). The gateway bearer token then travels over the network, so use it only for a host you control. n8nRemoteHost is empty by default.

  • getFilesClientModule, corpusClientModule and their RESEARCH_* environment variables name JavaScript that the Hub imports and runs with your user rights: set them only in configuration you control (see SECURITY.md.)

  • There is no application-level encryption of the local database; protect the data folder like any other research file. Do not send client-confidential material to providers. See SECURITY.md.

Limitations

  • Direct transport: offline-tested only (see the note at the top).

  • File-form and XML-archive application downloads (fileId/format) work only with the n8n transport.

  • Reference full-text search (research_reference_search), archiving and passage reading need the optional Get Files integration. Without it, references-fetch still downloads and hash-checks the official PDFs.

  • The hosted PatEx table on BigQuery dates from 2017; examiner outcomes after that need the official USPTO bulk release. Google Patents counts are publication counts, not filings or families.

  • The model proxy (src/model-proxy.mjs) and the PatentsView landscape form are n8n-only extras.

  • No licensed citator: later treatment of a decision is never checked.

  • Python packages in requirements.txt are version-pinned, not hash-pinned (the Windows installer installs wheels only, --only-binary=:all:, but does not verify hashes). The npm packages are pinned by package-lock.json with hashes.

  • PatentsView cohort intake (patentsview-intake, patentsview-install) accepts a corpus only if its normalized document names Context Forge as the normalizer (see docs/PATENTSVIEW.md); any other corpus client is refused.

  • Clean-machine installation on macOS and Linux has not been tested.

Testing

npm test        # offline: mocked providers, no network, no live services

Uninstall

Installed with Setup.cmd: use Patent Research Hub - Uninstall in the Start Menu. It stops Hub processes started from this installation (for example by an AI assistant), removes the program folder and shortcuts, and keeps the data folder unless you answer yes to deleting it. From a script: installer\remove.ps1 -KeepData -Yes or -RemoveData -Yes in the program folder. If the program folder or the data folder is itself a link (junction or symbolic link), only the link is removed and its target is left as it is; the output says so. Node.js and Python stay installed, and you remove the patent_research entry from your MCP client configuration yourself.

Installed from source: delete the install folder. Your data is in <dataDir> and <exportsDir> (inside the install folder unless you set PATENT_RESEARCH_DATA_DIR / PATENT_RESEARCH_EXPORTS_DIR); delete those too if you no longer need them, and remove the patent_research entry from your MCP client configuration.

More

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      A
      maintenance
      Exposes 20+ tools that let any MCP-compatible client run hybrid retrieval over MPEP/USC/CFR regulations, search 100M+ granted and published patents via Google BigQuery for prior art, and perform automated 35 USC 112 compliance checks on draft claims and applications. It also generates publication-style diagrams and flowcharts from natural-language descriptions.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables patent-rule lookup with citations over MPEP/USC/CFR via hybrid vector, lexical, and reranked search, plus prior-art searching across 100M+ patents through Google BigQuery. Also runs automated 35 USC 112 compliance checks on draft claims and applications and generates patent-style diagrams and flowcharts.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables AI clients to search patent rules and statutes with hybrid RAG, hunt prior art across 100M+ patents via BigQuery, run automated application compliance checks, and generate patent-style diagrams. It exposes 20+ tools that any MCP-compatible client can call for patent search, analysis, and figure creation.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Exposes 20+ tools that let any MCP-compatible client search patent rules and statutes (MPEP, USC, CFR) with hybrid RAG retrieval, hunt prior art across 100M+ patents via Google BigQuery, run automated claims and application compliance checks, and generate patent-style diagrams. It also provides CLI-driven setup, GPU-accelerated indexing, and health reporting so users can move from a raw invention description to a reviewable draft application.
      -