USPTO Patent & Trademark MCP Server

An MCP server (Python SDK 2.x, MCPServer) for accessing United States Patent and Trademark Office (USPTO) patent and trademark data through multiple APIs including the Patent Public Search API, the Open Data Portal (ODP) API, PTAB API v3, the TSDR trademark status API, and USPTO trademark search. Using this server, Claude Desktop can pull data from USPTO APIs, search through PTAB proceedings and decisions, research prosecution history, run trademark clearance searches, track trademark status, and more:

For an introduction to MCP servers see Introducing the Model Context Protocol.

Special thanks to Parker Hancock, author of the amazing Patent Client project, for blazing the trail to understanding of the string of requests and responses needed to pull data through the Public Search API.

Features

This server provides 42 tools across six live USPTO data sources (23 more, for APIs USPTO has shut down, stay in the code behind ENABLE_LEGACY_TOOLS):

1. Patent Search - Full-text search of granted patents and published applications via PPUBS

2. Full Text Documents - Get the text of a patent by number, or just its claims, abstract or front page to save context

3. PDF Downloads - Download patents as PDF files (Claude Desktop doesn't support this as a client currently)

4. Prosecution History - Transactions, file-wrapper document lists filtered by type (office actions, responses, notices of allowance), and the documents themselves as plain text or PDF via ODP

5. Office Action Search - Search 19.9M office actions (2007 on) by examiner language, rejection statute, art unit and date, with counts by document code and statute

6. Claim-Level Rejections and Citations - Which claims each office action rejected, objected to or allowed and under which statute, with Alice/Mayo/Myriad/Bilski flags (2012 on), and the references each action cited, searchable in reverse to find every application whose examiner cited a given patent

7. Petition Decisions - Search 1.1M decisions on patent petitions (revival, reinstatement, make special, supervisory review) by text, type, outcome and office

8. Patent Family Data - Continuity information, foreign priority, and related applications

9. Bulk Datasets - Search and access USPTO bulk data products including PatentsView disambiguated data

10. Forward Citations - Find the later patents that cite a patent

11. Trademark Search - Full-text search of US federal trademarks by mark text, owner, goods/services, and class (clearance/knockout searches)

12. Trademark Status & Documents - Authoritative live status, prosecution documents, and mark images via TSDR

13. Trademark Assignments - Recorded ownership transfer records from 1955 to present (no API key needed)

It runs locally over stdio for Claude Desktop and Claude Code, or over HTTP as a shared, stateless service — see Remote hosting over HTTP.

Note on unavailable APIs: The PatentsView API (search.patentsview.org) was shut down on March 20, 2026, with its data migrated to ODP bulk datasets. The Office Action and Enriched Citation APIs (developer.uspto.gov) were decommissioned in early 2026; office action search, rejections and citations have since returned on ODP (odp_search_office_actions, get_office_action_rejections, get_office_action_citations). The Patent Litigation API is not offered on the USPTO Open Data Portal; litigation data is available as a bulk download. The 23 affected tools stay in the code but are not registered by default, since their schemas cost every client about 6k tokens per session and they can only fail; set ENABLE_LEGACY_TOOLS=true to register them, and each returns workaround guidance pointing to a live tool.

Related MCP server: USPTO Patent MCP Server

API Sources

Prerequisites

• Python 3.10-3.13 (3.12 recommended); the server builds on the MCP Python SDK 2.x (mcp>=2,<3)

• Claude Desktop (for integration). Other models and MCP clients have not been tested.

• UV for Python version and dependency management

If you're a Python developer but still unfamiliar with uv, you're in for a treat. It's faster and easier than having a separate Python version manager (like pyenv) and setting up, activating, and maintaining virtual environments with venv and pip.

If you don't already have uv installed:

curl -LsSf https://astral.sh/uv/install.sh | sh

Installation

From PyPI (no clone needed)

uvx downloads and runs the published package on demand. Point Claude Desktop at it in claude_desktop_config.json, passing your API keys in the environment:

{
  "mcpServers": {
    "patents": {
      "command": "uvx",
      "args": ["patent-mcp-server"],
      "env": {
        "USPTO_API_KEY": "your_odp_key",
        "TSDR_API_KEY": "your_tsdr_key"
      }
    }
  }
}

or in Claude Code:

claude mcp add patents -e USPTO_API_KEY=your_odp_key -e TSDR_API_KEY=your_tsdr_key -- uvx patent-mcp-server

The keys are optional (see API Key Setup); the PPUBS patent tools, trademark search and assignment search work without any. A .env file is read from the working directory, which under uvx is wherever the client started the server, so pass keys through env as above.

From source

1. Clone this repository:

   git clone https://github.com/riemannzeta/patent_mcp_server
   cd patent_mcp_server

2. Install dependencies with uv:

   uv sync

3. Verify installation:

   uv run patent-mcp-server

   Should output:

   INFO     Starting USPTO Patent MCP server with stdio transport

API Key Setup

USPTO ODP API Key (Required for most tools)

To use the api.uspto.gov tools (ODP, PTAB), you need an Open Data Portal API key. Without it, these endpoints return 403 Forbidden. The Patent Litigation API is not offered on ODP and does not require an API key.

1. Create a USPTO.gov account at data.uspto.gov (requires ID.me verification)

2. Once signed in, visit "My ODP" in the site navigation to get your API key

3. See the Getting Started guide for detailed instructions

4. Complete your data.uspto.gov customer profile (job title, organization, intended use). ODP deactivates keys whose profile is incomplete, and a key that stops working with 401/403 usually means this.

5. Create a .env file in the patent_mcp_server directory:

   USPTO_API_KEY=your_actual_key_here

   Note: The PPUBS patent tools, trademark search (tm_* search tools), and trademark assignment search work without any API key.

TSDR API Key (Trademark status/document tools)

The TSDR tools (tsdr_*) require a TSDR-specific API key sent as the USPTO-API-KEY header. The ODP key does not work for TSDR — it passes the gateway's auth check, but every request then fails with BACKEND RESPONSE STATUS: 404.

1. Sign in to the USPTO API Key Manager with a free MyUSPTO account

2. Select the TSDR API product and click "Request API key" (the key is emailed and stored under your account)

3. Add it to your .env:

   TSDR_API_KEY=your_tsdr_key_here

TSDR rate limits (peak hours 5am-10pm ET): 60 requests/minute general, 4 requests/minute for PDF document bundles (120/12 off-peak). Use tsdr_list_trademark_documents (metadata only, not rate-limited like PDFs) before downloading bundles.

Trademark search and AWS WAF (no key needed)

tm_search_trademarks / tm_get_trademark use the internal API behind tmsearch.uspto.gov, which sits behind AWS WAF. It currently answers plain requests, but if USPTO tightens the WAF and searches start failing with 403/202 errors, copy the aws-waf-token cookie from a browser session on tmsearch.uspto.gov (valid ~4 days) and set:

TMSEARCH_WAF_TOKEN=your_cookie_value_here

Configuration

The server can be configured using environment variables in your .env file. All settings are optional with sensible defaults:

# API Keys
USPTO_API_KEY=your_key_here      # ODP/PTAB tools
TSDR_API_KEY=your_tsdr_key_here  # TSDR trademark tools (separate key — see above)
TMSEARCH_WAF_TOKEN=...           # Optional - only if trademark search hits the WAF

# Tool registration
ENABLE_LEGACY_TOOLS=false  # Also register the 23 tools for shut-down APIs (they return API_UNAVAILABLE)

# Logging
LOG_LEVEL=INFO  # Options: DEBUG, INFO, WARNING, ERROR, CRITICAL

# MCP Transport (see "Remote hosting over HTTP" below)
MCP_TRANSPORT=stdio    # stdio (default) or streamable-http
MCP_HOST=127.0.0.1     # Bind address when serving over HTTP
MCP_PORT=8000          # Port when serving over HTTP
MCP_PATH=/mcp          # URL path of the MCP endpoint
MCP_STATELESS=true     # Keep no per-client state between HTTP requests
MCP_JSON_RESPONSE=false # Reply with plain JSON instead of an SSE stream

# HTTP Settings
REQUEST_TIMEOUT=30.0  # Request timeout in seconds
MAX_RETRIES=3         # Maximum number of retry attempts
RETRY_MIN_WAIT=2      # Minimum wait time between retries (seconds)
RETRY_MAX_WAIT=10     # Maximum wait time between retries (seconds)

# Session Management
SESSION_EXPIRY_MINUTES=30  # How long to cache ppubs sessions
ENABLE_CACHING=true        # Enable/disable session caching

# API Endpoints (usually don't need to change)
PPUBS_BASE_URL=https://ppubs.uspto.gov
API_BASE_URL=https://api.uspto.gov          # ODP API endpoint (NOT data.uspto.gov)
TSDR_BASE_URL=https://tsdrapi.uspto.gov/ts/cd
TMSEARCH_BASE_URL=https://tmsearch.uspto.gov
TM_ASSIGNMENT_BASE_URL=https://assignmentcenter.uspto.gov
CPC_SCHEME_BASE_URL=https://www.uspto.gov/web/patents/classification/cpc/html

Claude Desktop Configuration

To integrate this MCP server with Claude Desktop:

1. Update your Claude Desktop configuration file (claude_desktop_config.json):

   {
     "mcpServers": {
       "patents": {
         "command": "uv",
         "args": [
           "--directory",
           "/Users/username/patent_mcp_server",
           "run",
           "patent-mcp-server"
         ]
       }
     }
   }

   You can find claude_desktop_config.json on a Mac by opening the Claude Desktop app, opening Settings (from the Claude menu or by Command + ' on the keyboard), clicking "Developer" in the sidebar, and "Edit Config."

2. Replace /Users/username/patent_mcp_server with the actual path to your patent_mcp_server directory.

When integrated with Claude Desktop, the server will be automatically started when needed and doesn't need to be run separately.

Claude Code Configuration

To integrate this MCP server with Claude Code for a particular project, from the project root:

claude mcp add-json patents '{"command": "uv", "args": ["--directory", "/path/to/patent_mcp_server", "run", "patent-mcp-server"]}'

If you're already running Claude Code, you'll have to /exit and restart. Then /mcp to verify that it's configured.

Remote hosting over HTTP

The server speaks stdio by default, which is what the Claude Desktop and Claude Code configurations above launch. It can instead serve the MCP endpoint over HTTP, so one deployment can back a whole team rather than every person running their own copy with their own API keys:

patent-mcp-server --transport streamable-http --host 0.0.0.0 --port 8000

The endpoint is then at http://<host>:8000/mcp. Point an MCP client at that URL — in Claude Code:

claude mcp add --transport http patents http://your-host:8000/mcp

This endpoint has no authentication of its own, and it holds your USPTO and TSDR API keys. Anyone who can reach it can spend your USPTO rate limits. Bind to 127.0.0.1 (the default) and put an authenticating reverse proxy in front of it, or otherwise restrict network access. The server logs a warning at startup when it is bound to anything other than loopback.

Requests are stateless by default: the server keeps no per-client state between them, so it can run behind a load balancer with several replicas and no session affinity. Each worker holds its own upstream USPTO Public Search session, which it establishes on demand and refreshes when it expires.

Each flag has a matching environment variable (see Configuration above); the command line wins where both are set.

Available Tools

Utility Tools

Patent Public Search (ppubs.uspto.gov)

Open Data Portal (api.uspto.gov)

PTAB Trial API (api.uspto.gov ODP v3.0)

TSDR - Trademark Status and Document Retrieval (tsdrapi.uspto.gov)

Trademark Search & Assignments

Note: tm_search_trademarks and tm_get_trademark use the undocumented internal API behind tmsearch.uspto.gov (the TESS replacement) — the same situation as the PPUBS patent search API. USPTO offers no official REST API for full-text trademark search. The request/response contract was verified live on 2026-06-10, but these tools may break without notice if USPTO changes the internal API. TTAB proceedings (oppositions/cancellations) have no REST API; daily TTAB XML is available as bulk datasets via odp_search_datasets.

Legacy tools (hidden unless ENABLE_LEGACY_TOOLS=true)

The 23 tools below are for APIs USPTO has shut down. They stay in the code and keep their names, but the server registers them only when ENABLE_LEGACY_TOOLS=true, because their schemas cost every client about 6k tokens per session and every call returns API_UNAVAILABLE. Each response carries a workaround naming the live tool that covers the need.

Patent Litigation API (Unavailable — not offered on ODP, issue #16)

All 4 Litigation tools return API_UNAVAILABLE. The Patent Litigation API is not listed in the ODP Swagger catalog. The OCE Patent Litigation dataset (74,000+ district court cases) is distributed as a bulk download at https://www.uspto.gov/ip-policy/economic-research/research-datasets/patent-litigation-docket-reports-data.

PatentsView API (Unavailable — shut down March 2026)

All 14 PatentsView tools return API_UNAVAILABLE with workaround guidance. PatentsView data has been migrated to the USPTO Open Data Portal as bulk downloadable datasets. Use ppubs_search_patents for patent search, odp_search_datasets to find bulk datasets.

Office Action APIs (Unavailable — decommissioned early 2026)

These 2 Office Action tools return API_UNAVAILABLE; odp_search_office_actions replaces them. get_office_action_citations and get_office_action_rejections are live again on ODP (v1.8.0).

Enriched Citation APIs (Unavailable — decommissioned early 2026)

All 3 Enriched Citation tools return API_UNAVAILABLE. Forward citations come from ppubs_get_citing_patents; the references a patent cites are on its front page (ppubs_get_patent_by_number(sections=["biblio"])).

Resources and Prompts

The server also provides MCP Resources (accessible via @ mentions):

• patents://cpc/{code} - CPC classification information at any depth (groups from USPTO's scheme pages)

• patents://status-codes - USPTO status code definitions

• patents://sources - Data source information

• patents://search-syntax - Query syntax guide (patents and trademarks)

• trademarks://classes - Nice/international trademark classes (1-45)

• trademarks://status-codes - Trademark status code definitions

And MCP Prompts (workflow templates). Each takes optional arguments; whatever you give is placed in a Subject section ahead of the steps, so the workflow starts with your specifics instead of asking for them:

• prior_art_search (invention, keywords, cpc_codes) - Comprehensive prior art search guide

• patent_validity_analysis (patent_number) - Patent validity analysis workflow

• competitor_portfolio_analysis (company, technology) - Competitor portfolio analysis

• ptab_proceeding_research (patent_number, proceeding_number, party) - PTAB proceeding research guide

• freedom_to_operate (product, cpc_codes) - FTO analysis workflow

• patent_landscape (technology, cpc_codes) - Technology landscape mapping

• trademark_clearance_search (mark, goods_services, classes) - Trademark clearance/knockout search guide

• trademark_portfolio_review (owner) - Trademark portfolio and deadline review

• trademark_status_monitoring (serial_numbers, mark) - Trademark status and conflict watching

Testing

The project includes comprehensive test suites:

# Run unit tests (default - skips integration tests)
uv run pytest

# Run with verbose output
uv run pytest -v

# Run integration tests (requires network access)
uv run pytest -m integration

# Run all tests including integration
uv run pytest -m ""

# Run with coverage report
uv run pytest --cov=patent_mcp_server

Integration tests save each tool's output to test/test_results/ for inspection. The directory is git-ignored: the files are rewritten on every run and nothing reads them back. Recorded responses that tests depend on belong in test/fixtures/.

The unit suite also runs in CI on every push to main and every pull request, across Python 3.10–3.13 (.github/workflows/tests.yml). Integration tests stay deselected there, so CI needs no API keys.

Development

To install development dependencies:

uv sync --dev

Publishing to PyPI

Releases go out through the Publish to PyPI GitHub Actions workflow (.github/workflows/publish.yml), which uploads with PyPI trusted publishing, so no API token is stored anywhere. After the version bump (pyproject.toml and the USER_AGENT in config.py) is merged to main, run the workflow from the Actions tab. It refuses a version PyPI already has, runs the tests, builds, publishes, and creates the vX.Y.Z tag and GitHub release.

To build locally:

rm -rf dist/ && uv run python -m build

Contributing

Issues and PRs welcome. See CONTRIBUTING.md for the contribution guide, and AGENTS.md for guidance specific to AI agents. Use the bug report or feature request templates when filing an issue — they prompt for the tool call, the constructed request URL/body, and the raw API response, which is usually enough to land a fix in one turn.

Version History

v1.8.3 (Current)

• Office action paging skips nothing after truncation: get_office_action_rejections pages by dataset rows but returns office actions, so when the token budget dropped actions, metadata.next_offset still pointed past their rows and the next page never showed them. It now points at the first row of the first dropped action, and rows_returned counts only the rows kept. get_office_action_citations had the same gap row for row and gets the same fix.

v1.8.2

• Keep truncated search results reachable: when the token budget removes records from the final page, has_more remains true. Continue at offset + count to retrieve the omitted records. Pages shortened only by stripping heavy fields keep their original pagination state.

v1.8.1

• Releases publish from GitHub Actions: the new Publish to PyPI workflow uploads through PyPI trusted publishing, so no API token is stored anywhere, and it tags each release and creates its GitHub release. This is the first version it published

• Docs: the legacy tool count now reads 23 everywhere (it said 25 in a few places after the two office action dataset tools went live in v1.8.0)

v1.8.0

• get_office_action_rejections and get_office_action_citations are live again, on ODP's claim-level datasets (/api/v1/patent/oa/oa_{rejections,citations}/v2/records) instead of the decommissioned developer.uspto.gov paths. Rejections come grouped by office action: the statutes it applies (101, 102, 103, 112, double patenting), the Alice, Mayo, Myriad and Bilski flags, and which claims were rejected, objected to or allowed. Filter by application or patent number, mail date, date range, statute, case law, document code and art unit, so case_law="alice", art_unit="3689" finds an art unit's Alice rejections. Citations list each reference an action cited and the section it was applied under; cited_reference runs the lookup in reverse

• Both tools return each action's document_id, which odp_download_document(format="text") reads

• Coverage: rejections from 2012 on (93M rows), about a week behind the mail date

• Tool count: 42 registered by default (65 with ENABLE_LEGACY_TOOLS)

• Verified live 2026-10-15

v1.7.0

• Full text of file-wrapper papers: odp_download_document takes format="text" and returns a paper's body as plain text from the XML USPTO publishes beside the PDF, so an office action (typically 10–20k characters) no longer has to be read as a base64 PDF. It covers office actions, notices of allowance, petition decisions, and applicant papers such as remarks (REM), claims (CLM) and IDS forms; amended claims keep their markup as ~~deleted~~ and ++inserted++. Long papers come in 24k-character pages split at paragraph breaks (text_offset / next_offset). A document without XML returns NO_TEXT_AVAILABLE, and the PDF remains the default format

• Parsed with the standard library; XML that declares entities is refused

• Verified live 2026-10-14 on CTNF, CTFR, NOA, REM, CLM, IDS, PET.PCT, PETDEC and SPEC papers from 2011 to 2026

v1.6.0

• New odp_search_office_actions searches ODP's Office Action dataset (19.9M actions mailed since mid-2007) by full text, document code, rejection statute (101/102/103/112/double patenting), art unit, date and patent number. Every response carries facet counts by document code and statute for the whole match set, so limit=1 answers counting questions. The API returns only the first 500 characters of each action; a hit's document_id is the file-wrapper documentIdentifier, so odp_download_document reads the rest

• New odp_search_petition_decisions searches 1.1M petition decisions (1991 on) by text, petition type code or words, outcome, deciding office, technology center, date and final-agency-action flag, with outcome, office and type facets. Query matches in the excerpt are marked «like this»; include_full_text adds each decision's full OCR text

• The get_office_action_text and search_office_actions legacy stubs now point to the new tool; check_api_status lists both datasets as active

• 401/403 from these endpoints carry a hint about ODP's customer-profile requirement

• Tool count: 40 registered by default (65 with ENABLE_LEGACY_TOOLS)

• Verified live 2026-10-14

v1.5.0

• CPC titles below subclass level: get_cpc_info (and patents://cpc/{code}) used a bundled table that stopped at the class level, so G06N 3/08 came back as "Computing; Calculating; Counting". Subclasses, main groups and subgroups now come from USPTO's published CPC scheme page for the subclass (www.uspto.gov/web/patents/classification/cpc/html/cpc-G06N.html; no key), fetched once a day per subclass: G06N 3/08 → "Learning methods", with the hierarchy G06N › 3/00 › 3/02 › 3/08 and the subgroups beneath it. Accepts every form the other tools print (G06N3/08, G06N 3/08, H04B7/0417 20130101). A code the scheme does not contain returns NOT_FOUND with the subclass's main groups; if the page cannot be fetched, the static section and class titles come back with a scheme_error

• Prompt arguments: all nine workflow prompts take optional arguments (patent number, company, mark, CPC codes, …). Whatever you give appears in a Subject section ahead of the steps; with no arguments the prompts read as before

• New CPC_SCHEME_BASE_URL setting; new cpc_scheme entry in check_api_status

• Verified live 2026-09-26: G06N 3/08, G06N, A61K, H04B 7/0417, and an unknown subclass and group

v1.4.0

• httpx → httpx2 for the nine USPTO clients and the logging transport. httpx2 is the successor from the same author (now under the pydantic organization) and is what mcp 2 already depends on; httpx 0.28.1 (December 2024) and httpcore 1.0.9 (April 2025) have stopped moving. The API is the same for everything this code uses (AsyncClient, AsyncHTTPTransport, Cookies, build_request/send(stream=True), the exception classes), so the change is a rename across 20 files; httpx and httpcore leave the dependency tree entirely. The PPUBS session handling, PDF print job, TSDR XML document list, tmsearch, Assignment Center v3 and the ODP document download redirect were each re-verified live on 2026-09-26 (whole suite, nothing deselected)

v1.3.0

• MCP Python SDK 2.x: the server now builds on mcp>=2,<3 (MCPServer, the renamed FastMCP), which lifts the mcp<2 pin v1.2.1 added. Tool names, arguments and results are unchanged; the --host/--port/--path/--stateful/--json-response flags and MCP_* variables keep their meaning. Under the hood the transport settings move from the server constructor to run_streamable_http_async(), the protocol tests use mcp 2's in-memory Client, and the dependency drops the [cli] extra (empty in 2.x)

• Structured output off: mcp 2 sends every dict result twice by default, as JSON text and again as structured_content, against an output schema of "any object" — measured at 2.0x wire size on documents and searches. Tools register with structured_output=False, so results travel once, as before

• Pinned click>=8.3.3 (PYSEC-2026-2132; a transitive dependency via uvicorn that predates this release) so pip-audit is clean on the new tree

• Verified 2026-09-26: unit suite on mcp 2.2.0; the live suite (unit plus integration, nothing deselected); a clean uvx install from the built wheel; stdio and streamable-http smoke tests through mcp 2's own client

v1.2.1

• Fixed fresh installs: mcp[cli]>=1.27 resolved to mcp 2.x, which renamed mcp.server.fastmcp to mcp.server.mcpserver, so uvx patent-mcp-server and pip install patent-mcp-server crashed on import (ModuleNotFoundError: No module named 'mcp.server.fastmcp'). mcp 2.0.0 shipped on 2026-07-28, so v1.1.1 and v1.2.0 were both affected; the checked-in uv.lock (mcp 1.28.1) hid it from local runs and CI. The dependency is now mcp[cli]>=1.27,<2. Migrating to the mcp 2 API is a separate piece of work

v1.2.0

• Fixed patent-number parsing: "US 9,876,543 B2" was read as 98765432 (the kind code folded into the number) and "D845123" as 845123, a 1907 utility patent. Numbers now keep their D/RE/PP prefix and lose separators, a leading "US" and the kind code. The lookup also dropped a dead patentNumber:"…" query that returned nothing and cost ~3 s per call

• sections on ppubs_get_full_document and ppubs_get_patent_by_number: choose biblio, abstract, claims, description. A whole document is ~24k tokens; ["biblio", "claims"] is ~6k. Empty and search-highlight fields are dropped from every document

• Smaller search results: each PPUBS hit carried two 7 KB references-cited lists; three hits are now 5k characters instead of 41k. ppubs_search_* default to limit=20 (the truncation step cut larger responses to 20 anyway)

• New odp_download_document fetches a file-wrapper document (office action, response, notice of allowance…) as PDF, following ODP's signed 30-second redirect; documents over 4 MB are refused. odp_get_documents gains document_code and direction filters, paging, and a per-code count of the whole wrapper

• New ppubs_get_citing_patents: forward citations via the .urpn. field of granted patents; total is the citation count

• Legacy tools hidden by default: the 25 tools for shut-down APIs made up 39% of the schema text sent to every client (~6k tokens). Set ENABLE_LEGACY_TOOLS=true to register them; the functions, names and workaround messages are unchanged

• Read-only annotations on every tool, so clients that honor them need not confirm each call

• The competitor-portfolio, freedom-to-operate and patent-landscape prompts, the patentsview_* workaround messages and this README still taught the slash-prefix search syntax that stopped working in August (AN/, IN/, CPC/); all now use .as., .in., .cpc. and @pd

• Removed test/test_patents.py: two script-style live checks that logged errors instead of asserting, so they always passed while calling USPTO and rewriting json/ and pdfs/ on every default pytest run, including CI. The default suite now makes no network calls

• Fixed tm_search_assignments: Assignment Center moved its search API from /ipas/search/api/v2/ to /v3/, and CloudFront now refuses POST on the old path with a 403 HTML page. Same request body and response envelope; only the path changed

• 429 handling in the ODP and PTAB clients: a rate-limited call now waits (Retry-After, else exponential backoff) and retries up to MAX_RETRIES instead of returning the 429 to the caller. api.uspto.gov limits per key, so a burst of tool calls tripped it

• Integration tests: test/test_tools.py and test/test_tools_pytest.py were duplicates; the pytest one survives under the shorter name, minus eight skipped PatentsView tests, with its patentNumber:"…" queries and numFound assertions (neither matched the live API or the tool envelope any more) fixed, and with live tests for sections, prefixed numbers, ppubs_get_citing_patents, the odp_get_documents filter and odp_download_document. uv run pytest -m "" runs the whole suite, unit and live, and passes

• Live-verified 2026-09-26: .pn. with D/RE prefixes, .urpn., the ODP document listing and download redirect, Assignment Center v3

• Tool count: 38 registered by default (63 with ENABLE_LEGACY_TOOLS)

v1.1.1

• Fixed ppubs_download_patent_pdf: USPTO moved the PDF download endpoint — the old /api/internal/print/save/{pdfName} path now returns 404; the client uses the live /api/print/save/{pdfName} endpoint (verified live 2026-08-05 with a real search, document fetch, and PDF download)

• Corrected PPUBS search syntax guidance: the slash-prefix field qualifiers (TTL/, IN/, AN/, CPC/) no longer work on the live API — they silently return 0 results, and TTL/"phrase" returns a server 500. Search tool docstrings, the prior-art prompt, and the patents://search-syntax guide now teach the working dotted-suffix forms (.ti., .ab., .clm., .spec., .in., .as., .pn., .cpc., @pd/@ad date ranges), each verified live

• Both breakages were USPTO-side drift predating v1.1.0 (confirmed by running the same live test against v1.0.0-era code)

v1.1.0

• Remote hosting over HTTP: new --transport streamable-http mode serves the MCP endpoint over the network, so one deployment can back a whole team instead of every user running a local copy. stdio remains the default, so existing Claude Desktop and Claude Code configurations are unchanged. New flags --host, --port, --path, --stateful, --json-response, each with a matching MCP_* environment variable. See "Remote hosting over HTTP" for the security caveat — the endpoint holds your API keys and does not authenticate callers

• Stateless by default: HTTP requests carry no per-client state, so the server can run behind a load balancer with several replicas and no session affinity

• Fixed a session race in the Public Search client (affects stdio users too): concurrent tool calls each reset the shared cookie jar and raced to swap the access token, so requests already in flight could be signed with a half-replaced session. Session setup is now serialized, a 403 refresh is skipped when another call has already replaced the token, and the token travels per request instead of living on the shared client's default headers

• Fixed shutdown: closing the nine HTTP clients no longer happens in an atexit hook that spun up a fresh event loop; it now runs in the loop the clients were opened on. Deliberately not a FastMCP lifespan — in stateless HTTP mode that runs once per request, which would close the clients after the first tool call

• Raised the MCP SDK floor to >=1.27 (was >=1.3.0, which allowed installs to resolve an SDK without streamable HTTP)

• Added MCP protocol-layer tests: the suite previously called tool functions directly and never exercised schema generation, resource/prompt registration, or serialization. 19 new tests (378 total, up from 359)

v1.0.0

• Trademark support: 9 new trademark tools across three new clients, all verified against the live USPTO services on 2026-06-10

  • TSDR (tsdr_get_trademark_status, tsdr_list_trademark_documents, tsdr_download_trademark_documents, tsdr_get_trademark_image) — official trademark status/document API. Requires a TSDR-specific key (the ODP key does not work); error responses detect the wrong-key signature and explain how to get the right one. Document bundles above 4 MB are rejected with filter guidance (full wrappers can exceed 10 MB)

  • Trademark search (tm_search_trademarks, tm_get_trademark) — full-text search by mark text, owner, goods/services, and Nice class via the internal Elasticsearch API behind tmsearch.uspto.gov (no official REST API exists). Verified live; handles AWS WAF rejections with TMSEARCH_WAF_TOKEN support

  • Trademark assignments (tm_search_assignments) — USPTO Assignment Center public API (assignmentcenter.uspto.gov, no key required), searchable by serial/registration number, assignee, assignor, and reel/frame. Replaced the legacy assignment-api.uspto.gov XML API decommissioned June 5, 2026

• New reference tools and resources: get_trademark_class_info, get_trademark_status_code, trademarks://classes, trademarks://status-codes (all 45 Nice classes, common trademark status codes)

• 3 new workflow prompts: trademark_clearance_search, trademark_portfolio_review, trademark_status_monitoring

• Fixed ppubs_download_patent_pdf (called download_image with the wrong signature, raising TypeError)

• Rewrote patent workflow prompts to reference live tools (the old prompts still pointed at decommissioned PatentsView/Office Action/citation tools)

• New env vars: TSDR_API_KEY, TMSEARCH_WAF_TOKEN, TSDR_BASE_URL, TMSEARCH_BASE_URL, TM_ASSIGNMENT_BASE_URL

• Tool count: 61 registered (36 active, 25 unavailable)

v0.9.5

• Re-enable 7 PTAB tools on USPTO ODP v3.0: ptab_search_proceedings, ptab_get_proceeding, ptab_get_documents, ptab_search_decisions, ptab_get_decision, ptab_search_appeals, ptab_get_appeal (issue #23). PTAB data relocated to ODP /api/v1/patent/trials/* and /api/v1/patent/appeals/* (paths not in the ODP Swagger UI); the standalone-API decommission (issue #16) was correct for the Patent Litigation API, but PTAB moved rather than disappeared.

• Active tool count: 27 (up from 20); unavailable: 25 (down from 32); total registered remains 52

v0.9.4

• Fix ppubs_search_patents / ppubs_search_applications query semantics (issue #21): default operator changed from OR to AND, so multi-word queries like machine learning no longer match the entire corpus and collapse into the latest-grants fallback under date_publ desc sort.

• Fix template-mutation bug in PPUBS client (search_query.copy() → copy.deepcopy(...)), eliminating a concurrency hazard between parallel calls.

• Fix odp_search_applications filters being silently ignored upstream (issue #21): switched from GET query-string params to POST with a Lucene-style q body. assignee_name, inventor_name, application_number, patent_number, and filing-date ranges are now properly AND-combined into the search. Tool now returns MISSING_FILTER rather than dumping the full 12.8M-record corpus when called with no filters.

• Updated ppubs_search_patents / ppubs_search_applications / odp_search_applications docstrings to reflect the corrected semantics and document Lucene query support on ODP.

• Added CONTRIBUTING.md, AGENTS.md, bug-report + feature-request issue templates, and a PR template.

v0.9.0

• Handle PTAB Trial API and Patent Litigation API unavailability on ODP (issue #16)

• All 7 ptab_* tools and 4 litigation tools now return API_UNAVAILABLE with workaround guidance pointing to PPUBS tools and USPTO bulk datasets

• Active tool count: 20 (down from 31); unavailable: 32 (up from 21); total registered remains 52

• Added unit tests for all 11 newly-unavailable tools and extended the shared error-structure parametrization

• Updated check_api_status, resources.py data sources, client docstrings, and README to reflect the shutdown

v0.8.0

• Handle decommissioned PatentsView API (shut down March 20, 2026)

• All 14 patentsview_* tools return API_UNAVAILABLE with workaround guidance

• Fixed circular references in office_actions resources that pointed to unavailable PatentsView tools

• Updated API Sources table, configuration, and documentation

v0.7.0

• Handle decommissioned Office Action and Enriched Citation APIs (developer.uspto.gov)

• All 7 affected tools return API_UNAVAILABLE with workaround guidance

• Added test/unit/test_unavailable_tools.py for decommissioned tool testing

• Code cleanup: removed dead code, improved docstrings

v0.6.2

• Updated API key registration instructions: keys are now obtained from data.uspto.gov ("My ODP")

• Clarified that api.uspto.gov is the correct API endpoint (not data.uspto.gov which is the web portal)

• Noted PTAB API v3 migration to ODP and Office Action API migration (early 2026)

v0.6.1

• Added PatentsView attorney search tools (patentsview_search_attorneys, patentsview_get_attorney)

• Added PatentsView IPC classification tools (patentsview_lookup_ipc, patentsview_search_by_ipc)

• Fixed bug in search_publications method (pagination options not being passed)

v0.6.0

• PyPI release preparation

v0.5.0

• Focused on USPTO-only data sources

• Renamed ODP tools with odp_ prefix for clarity

• Improved function signatures (using query instead of q)

v0.3.0

• Added 33 new tools (PTAB, PatentsView, Office Actions, Citations, Litigation)

• Rate limiting support for PatentsView API

• Comprehensive async client architecture

v0.2.2

• Centralized configuration with environment variables

• Standardized error handling

• Input validation with Pydantic

• Retry logic with exponential backoff

• Session caching for PPUBS

License

MIT