PatentRef

Server MCP untuk PatentRef: pohon klaim paten AS, ringkasan, dan pencarian melalui API yang kompatibel dengan PatentsView

GitHub

Dokumentasi

patentref-mcp

patentref-mcp MCP server - quality and maintenance score on Glama

patentref-mcp is an MCP server for the PatentRef API (https://patentref.io), a PatentsView-compatible API for US patents. Its 15 read-only tools give an agent a US patent's claim tree with element segmentation, antecedent links and character offsets (built by deterministic code), a patent brief that fits a token budget, the expiry and maintenance-fee status record, search over US grants from 1976 to now, patents similar to a text, and patent office rules with their official sources.

How it runs

The server runs on your own machine. Your MCP client starts it as a local process and talks to it over stdio, the default. Each tool call becomes one HTTPS request from your machine to the PatentRef API (https://patentref.io) with the key from your environment, and the tool returns the API's JSON as the API sent it. There is no hosted MCP endpoint for this server: you run it yourself.

patentref-mcp --http is for MCP clients that connect to a URL instead of starting a process. It serves the same tools as streamable HTTP from your machine, at http://127.0.0.1:8787/mcp by default (--host and --port change the address). The requests to the API still go out from your machine with your key, so anyone who can reach that address can use your key's allowance: keep the default 127.0.0.1 unless you mean other machines to reach it.

Run

pip install patentref-mcp
export PATENTREF_API_KEY=...        # a key from https://patentref.io/api/v1/meta/signup/
patentref-mcp                       # stdio, what most MCP clients start
patentref-mcp --http --port 8787    # streamable HTTP at http://127.0.0.1:8787/mcp

Claude Desktop (claude_desktop_config.json):

{"mcpServers": {"patentref": {"command": "patentref-mcp", "env": {"PATENTREF_API_KEY": "..."}}}}

Environment: PATENTREF_API_KEY (sent by every tool but build_facts and search_disclosures), PATENTREF_BASE (default https://patentref.io), PATENTREF_TIMEOUT (seconds, default 30).

A free key allows 1,000 requests a month and 45 a minute. Each call of a keyed tool is one request; answers 400 and 404 count too, 403 and 429 do not. build_facts and search_disclosures send no key.

Check an install: python -m patentref_mcp.smoke 10000364 starts the server over stdio, lists its tools, makes two tool calls and prints the transcript. patentref-mcp --version prints the version.

Tools

  • check_patent_status: Get the expiry and maintenance-fee status record of one granted US patent: status (active, expired_term, expired_fee_nonpayment, withdrawn or unknown), its basis, reinstatement_possible with its legal basis, as_of, data_version, a versioned permalink, a one-sentence plain statement and next_steps. Record facts, not a legal opinion. Use it for one known patent number; use search_patents with the "lapse" field group for the status of many patents at once, and get_patent_brief for the claims together with the status.

  • search_patents: Search granted US patents by field values with the PatentSearch v1 query grammar and return the matching rows, a page at a time. Use it to find patents by title words, dates, CPC class, assignee, inventor or other fields; use similar_patents to rank patents by closeness to a passage of text, check_patent_status for one known number, and search_disclosures for PatentRef registry records rather than patents. Query operators: _eq, _neq, _gt, _gte, _lt, _lte, _begins, _contains, _text_all, _text_any, _text_phrase, _and, _or, _not; dotted names reach nested fields. The "lapse" field group adds expiry_status, expiry_date_estimated, expiry_basis, reinstatement_possible, fee_status and lapse_date. For the next page, pass the last row's sort values as after. Field names and the full grammar: https://patentref.io/llms-full.txt.

  • similar_patents: Rank granted US patents by how close their title and abstract are to a passage of text, and return the n nearest with data_version and as_of. Use it when you have a description, an abstract or a claim and want related patents; use search_patents for exact filters such as dates, classes or names. The ranking is full-text bm25 over titles and abstracts, searched with the rarest words of the text, so the answer is a record of what the data holds, not a search opinion.

  • get_claim_tree: Get every claim of one granted US patent as a tree parsed from the claim text the USPTO published: per claim its number, category, dependency, preamble, transition, elements, term mentions (each "the" or "said" phrase linked to its earlier mention), claim references and flags, with structure_version, data_version, text_data_version and as_of. Every part carries start and end offsets into its claim's text (text[start:end] is the part). Use it when you need all claims in full detail; use get_patent_brief for a shorter answer that fits a token budget, and check_patent_status for the status alone. Flags are notes on the text and the data, not conclusions about the claims. A patent without claim text in the data answers 404.

  • get_patent_brief: Get a brief of one granted US patent that fits a token budget, in a fixed order: identity and dates, the status record with its basis, each independent claim as its elements with start and end offsets into the claim text, one line per dependent claim, and the claim term keys. The brief stops at the item that would go over max_tokens; truncated and omitted say what was left out, and tokens gives its size, never over max_tokens. Every claim and element carries an anchor: page + "#" + anchor opens it on the record page. Quotes are the claim text as published. Use it to read or summarize a patent within a context budget; use get_claim_tree for all claims in full structure, and check_patent_status for the status alone.

  • rep_required: List the representation rules of a patent office: whether and when an applicant must act through a registered agent, each row with its official source URL and a verbatim quote. Use it for questions about needing a local agent; use fee_schedule for fees, route_windows for time limits, and rules_index for the offices and the act vocabulary. office takes US, EP, CN, JP, KR, AU or PCT.

  • fee_schedule: List the official fees of a patent office in force on a date: filing, search, examination, grant, renewal and other fees with their amounts, currency and entity tiers as the cited schedule prints them, each row with its official source URL and a verbatim quote. A row is in force when effective_from <= as_of < effective_to. Use it for what an office charges; use rep_required for agent rules and route_windows for time limits. office takes US, EP, CN, JP, KR, AU or PCT.

  • route_windows: List the generic time windows of a patent office, such as the PCT national phase time limit in months from the priority date, each row with its official source URL and a verbatim quote. Use it to learn how long a period an office allows; it never computes a last day for a specific application and takes no application number or date. Use fee_schedule for fees and pph_routes for Patent Prosecution Highway routes. office takes US, EP, CN, JP, KR, AU or PCT.

  • pph_routes: List the Patent Prosecution Highway (PPH) rows of to_office whose cited source names from_office as a partner office, each with its official source URL and a verbatim quote, and to_office_pph_rows, the number of PPH rows held for to_office. Use it to see whether work done at one office can support a PPH request at another; use route_windows for time limits. An empty list means no PPH row of to_office names from_office.

  • ai_use_statement_required: List the rows on whether a patent office requires a statement on the use of AI, each with its official source URL and a verbatim quote. An empty answer means the data holds no cited rule for that office; nothing follows from its absence. Use it for AI-use statement rules; use rep_required, fee_schedule or grace_period for the other rules of an office. office takes US, EP, CN, JP, KR, AU or PCT.

  • grace_period: List what the cited law of a patent office says about disclosures made before filing (the grace period), each row with its official source URL and a verbatim quote. Citations only: it does not say whether any disclosure falls inside the period. Use it to read an office's grace period text; use search_disclosures to find dated public disclosures in the PatentRef registry, and route_windows for other time limits. office takes US, EP, CN, JP, KR, AU or PCT.

  • rules_changelog: List the releases of the patent office rules data from newest to oldest, each with its version, date and the rows it added, closed, changed or removed. Use it to see what changed since a version or date you used before; use rules_index for the current version, the offices and the vocabulary.

  • rules_index: Describe the patent office rules data: its version and as_of date, the offices it holds with their row counts by rule kind, each rules tool with its parameters, and the act vocabulary. Call it before the other rules tools to learn which offices and act values exist; use rules_changelog for the history of changes. Takes no arguments, so nothing can be filtered here.

  • build_facts: Get the facts of the patent data build the API serves now: data_version, as_of, refreshed_at, the snapshot hash, superseded versions and the table delta. Use it to cite or check which data an answer came from, or to see when the data was last refreshed; it holds no patent data (use search_patents or check_patent_status) and no rules data (use rules_index). Takes no arguments.

  • search_disclosures: Search the public records of the PatentRef registry (dated technical disclosures) whose title or abstract match q, from newest to oldest, up to 20. Each record carries its permalink (https://patentref.io/r/ and the record's id), server_time, bundle_hash and proof_status; the submitter's text is inside user_content, as entered. A record shows what was published and when, not a patent or an opinion. Use it to find disclosures; use search_patents or similar_patents to find patents.

Every tool is read only and carries the MCP annotations readOnlyHint, idempotentHint and openWorldHint. An error from the API comes back as the tool's error text: the HTTP status and the API's X-Status-Reason. The server does not retry.

Changes

See CHANGELOG.md in the repository: https://github.com/patentref/patentref-mcp/blob/main/CHANGELOG.md

License

MIT, see LICENSE. The data the API returns has its own license and attribution lines: https://patentref.io/LICENSES.md

Source: https://github.com/patentref/patentref-mcp. The package is published on PyPI (https://pypi.org/project/patentref-mcp/). Questions and problems: hello@patentref.io.