Seven servers · one subscription · one prosecution record
Seven MCP servers for Claude, and for any other platform that speaks the protocol. Patent search at the front door, free on its own and uncapped here. Behind it, the text of the law and the one that dates it, which with the search make USPTO Research Bundle, and the prosecution record, the Board's docket, the Office's petitions and the examiner's citations, which make USPTO Full Bundle.
Start with USPTO Full Bundle Follow the twelve-step walkthrough Read the case study
Before chapter one: USPTO Patent Search, the free front door, is part of this plan with its monthly cap lifted.
Prologue
None of these is a search that fails. Each one is a search that succeeds and returns something a practitioner has to catch by hand.
The examiner's reasoning is the thing worth reading, and for a long stretch of the record it exists as a scanned page inside a document bag rather than as text. A keyword index cannot see into it, so a rejection can be present in the file and absent from every search run over the file.
The Office publishes a field called the effective filing date. For a national-stage case it is the date of entry under section 371, and for a continuation it is the child's own filing date. Build a prior art cutoff on it and the cutoff inverts: references that are prior art stop looking like prior art, and the search returns cleanly.
Pulling the paper itself is the easy part. Naming it is not. What lands in the downloads folder is a PDF called after the identifier that fetched it, so an afternoon of research ends as a row of files that can only be told apart by opening them. The research succeeded. A week later the record of it is unusable.
The file wrapper, the Board and the petitions each answer with a download link that stays live for a day, and the file arrives named out of the record rather than out of the request: the application or the proceeding, the patent, and what the paper is.
Discretionary denial practice has been rewritten repeatedly since 2018, and the Office rescinded its own interim procedure on 28 February 2025. Ask about a petition filed in 2023 and the rescinded memorandum is the answer. Ask the same question without a date and you get today's regime applied to a proceeding it never governed.
A rule applied to the wrong side of a regime change is wrong silently. Nothing in the answer looks different.
Chapter I USPTO Patent Search
Every US patent grant since 1976 and every published application since 15 March 2001, searched in place: PatFT field codes, semantic search by meaning, keyword search by exact words, and the theme layer that names the neighbourhood a subject falls in. On its own it is the free tier, 150 calls a calendar month. Here the cap comes off.
Both paid tiers include the unlocked edition: searches without a monthly count, and the browser search page, laid out like the Office's former PatFT, which any results panel opens and the free tier does not get. Every result names the calls on the servers below that reach the live record of what it found, which is why the search comes first and the rest of this page follows it.
The search has a page of its own, with the real queries and what they returned. Read it there.
Chapter II MPEP retrieval
Search over legal text fails in two opposite directions, and both failures read as success. Keyword search finds the section that happens to use your term and misses the section that governs your situation under a different name. Meaning-based search finds the passage that is about your topic and cannot tell a term of art from a description of one, so it will hand you a discussion of restriction practice when you asked about a restriction requirement.
So this server runs both and fuses them, and the balance between them is a dial rather than a setting. It is turned per question, by the assistant making the call, which is the only party that knows whether what was just typed is a defined term or a theme. Reranking can be tightened, loosened or switched off for a single question without changing anything for the next one.
Both halves of that retrieval can be pointed at the firm's own machines. The embedding endpoint is configuration rather than architecture, so the model that turns a question into a vector can be one the firm hosts itself, and the reranker is the same: bring your own, or run a question with reranking switched off and take the fused ranking as it stands. The text being searched is published government record. The question asked of it is usually the work product, and the question is the part that never has to leave the building.
The corpus is stratified by legal weight, and the stratification is enforced as a filter rather than described in prose. Statute sits above binding case law. Below the courts, and beside the manual rather than above it, sit the Board's precedential and informative decisions, Director review decisions and the Commissioner's decisions. The manual itself sits above the Federal Register and Official Gazette notices that can supersede it, and all of that sits above examination guidance, which never supersedes anything. That order is what lets an answer say which of two passages wins.
A retrieved passage arrives knowing where it sits. The range of the section around it travels with the result, so reading the whole section is one more call rather than three. A chunk is the right size for finding a rule and frequently the wrong size for reading one, because the sentence you matched starts in the middle and the definition you need is in the paragraph above it.
A question can also be asked in several framings at once. Name a domain and the server runs that domain's own hand-written framings of the question in parallel and fuses them, on the principle that no single phrasing of a legal question is reliable and a set of them behaves better than the best one alone.
What is in it: the Manual of Patent Examining Procedure, Ninth Edition, Revision 01.2024, complete from chapter 0100 through 2900 with its appendices, more than two thousand eight hundred court opinions and orders running from 1810 to this year, most of them Federal Circuit and most of the rest Supreme Court, the Board's precedential and informative decisions and Director review decisions, Federal Register documents from 1995, and Office notices from 1994. All of it public record, all of it carrying the source it came from.
Two rules the corpus is built under. A source link is never written unless it was fetched and answered, so every passage the server returns carries the direct public address of the document it came from, the USPTO page, the court's own PDF, the Federal Register entry, and the model can hand you that address beside the sentence it supports, for checking the citation yourself. Where a source has only an index page, the passage says so and gives the citation as text rather than a guessed link. And text is never paraphrased on the way in, so what comes back out is what the document says.
On the summaries
Each document carries a machine-written orientation summary, and the server labels it as exactly that: an aid for deciding what to open, not evidence. It has been observed to get procedural direction wrong, saying a case was remanded where it was reversed. The label ships with the summary rather than living in documentation nobody reads, and the finding you cite comes from the opinion.
Chapter III The hypergraph
The division of labour between these first two chapters is the plainest thing about them. The text server holds the documents that apply and record the law. This one holds the documents that change it.
It keeps prosecution and post-grant facts as a temporal graph, and over that graph a register of framework shifts: Supreme Court and Federal Circuit decisions, Office guidance and rulemaking, designated Board decisions and the trials themselves, each carrying the date it took effect and, where one exists, the date something displaced it. A shift never edits the facts it re-governs. The prosecution record is evidence and stays as filed; the governance is a separate layer attached over the top of it.
Ask it a question and the date is required. There is no default to today, on purpose, because every failure this server exists to prevent is a question that silently used now. Each named pattern declares which event its date should be, so a petition question wants the petition filing date, a time-bar question wants the date the complaint was served, and an appeal question wants the notice of appeal.
What comes back is the regime sorted into four groups: in force on that date, not yet in effect on that date, already superseded by it, and, separately, any shift the corpus holds with no effective date at all. That fourth group exists because an undated rule silently filed under "not in force" is precisely the failure the tool was built to stop. No model is asked to judge any of it: the retrieval fuses its sub-queries by rank, and the dating is arithmetic over dates the corpus carries.
The semantic half of this server is configurable the way the text server's is. Point it at an embedding endpoint the firm runs, bring your own reranker or none, and the questions put to the graph stop at the firm's own boundary. The dating never left it in the first place: comparing two dates is arithmetic, and arithmetic does not call a vendor.
Ten of the framework shifts bearing on discretionary denial, read out of the server on 10 October 2026. The dates are its own. Set the date a petition was filed and the register sorts itself.
Petition filed June 2023
Nothing about a row changes when you move this. Its dates were fixed the day it issued. What changes is which of them your question falls between.
A petition filed in June 2023 is governed by a memorandum that no longer exists. A search that never asked for a date will not mention it.
That is the whole argument for dating a legal graph. The rescission is real and current and correct, and applying it to a 2023 petition is still wrong. So the server reports the two facts separately: what governed then, and what has displaced it since, with the direction of the displacement on the edge.
Proposals are handled the same way, which is to say they are not promoted. A rule that has only been proposed sits in the corpus as retrievable text and is not registered as a shift, so an as-of query can quote the proposal's language and never count it as law. The same discipline runs forward through the guidance line: the Office's February 2024 inventorship guidance for inventions made with the help of artificial intelligence, 89 FR 10043, carries its own supersession by the revised guidance at 90 FR 54636 on 28 November 2025, so a question dated before that November still gets the earlier one.
Retrieval is built to find what resembles. An attorney is paid to find what differs. A date is one of the places those two come apart.
That pair is worth staying with, because it is where searching by resemblance fails hardest and most quietly. Guidance that matters arrives with a programme behind it: the notice, then the memoranda to examiners, the training material, the outreach sessions and everything written about all of that. The short revision that displaces it two years later arrives with none of it. So the older regime is the one an index has more of, and volume is what a similarity score is counting when it says a passage is the closest match. Nothing about that ranking changes on the day the rule stops being law, because the rule stopping is not a fact about the text.
Which is why the order of operations here runs the other way round from the usual one. The graph settles what was law on the date in question first, by arithmetic over dated edges, and retrieval then works inside that perimeter and fetches the words. Neither half is asked to do the other's job: nothing is ranked into force, and nothing is dated by resemblance. The same discipline covers the other ways a legal document loses its meaning in transit. A passage keeps the section it came from, because the identical sentence under two different section headings can carry opposite consequences, and a clause keeps what it hangs off, because an exception read without the rule it modifies is not a weaker answer but an inverted one.
A structure built this way is not primarily a faster way to look things up. It is a defensive one. The point is that a fluent, well-cited, confidently worded answer cannot be assembled out of a rule that was withdrawn last year, and that when an answer does come back, what it rests on is inspectable: which fact, which edge, which date, and which parts of the response the server worked out rather than read. Speed is what a firm notices. The defence is what it is actually buying.
Where the authority that displaced a rule sits outside the part of the corpus a question reached, the answer says so in those words. There is a real difference between nothing having displaced a rule and not being able to see what did, and only one of them is a finding.
Most knowledge graphs store a fact as a subject, a predicate and an object. Prosecution facts rarely have two ends. One finding out of this corpus, on the Board declining to exercise discretion in IPR2020-01019, carries six participants and every one of them has a named role: the Board as the decisionmaker, patent RE47,353 E as the challenged patent, section 314(a) as the discretionary institution statute, the inter partes review as the proceeding, the district court as the parallel forum, and the proceeding number as the identifier.
Split that into pairs and the roles are the first thing lost. You are left knowing that a district court and a patent and a statute are associated with a proceeding, and not knowing which was the forum, which was challenged and which was applied. The same loss is worse in an obviousness rejection over two references, where which one is the primary reference and which one is secondary is a legal position and not a detail. Facts here carry three or more participants by construction; anything genuinely two-ended stays an ordinary edge, because that is what it is.
Every response names the fields it computed rather than read. Whether a shift was in force on your date is arithmetic the server did. The note explaining how a shift re-governs a fact is the server's characterization of it, not language quoted from the authority. Even a citation link is flagged, because it is built by rule from the source document rather than read off the record. All of it is useful and none of it is a source, so the labelling travels with the answer and nothing downstream inherits an inference as a finding.
The same care runs down to how a shift bears on a fact. It can confirm it, narrow it, expand it, clarify it or abrogate it, and it can also be recorded as bearing on it in none of those ways. That last one is a finding, not a blank: somebody looked and found no governance, which is different from nobody having looked yet, and the two are stored differently.
Everything in this corpus is published government record: the manual, the regulations, case law, examination guidance and Office guidance. A single answer reports which of those bodies attested a doctrine, so the same rule found in three of them arrives corroborated rather than counted three times.
Chapter IV Patent File Wrapper
The Patent File Wrapper server returns the examiner's actual text in a single call. Not a document list and not a document to go and read somewhere else: the body of the office action, or one section of it, as text you can quote. Ask for the section 103 rejection in the final action and that is what arrives.
Much of a file wrapper is not text at all. Papers were filed as images, and a flat scanned PDF is a picture of a document as far as any index is concerned. So there is a second way to read, for the documents the first way does not reach: the server fetches the document, takes the text where the text exists, and where the page is an image reads it with OCR. What comes back is the words on the page rather than the fact that a page exists, which is the difference between a model that can advise on a document and one that can only report that it is there.
The two are worth telling apart. The indexed lane is one call and it answers for office actions. The reading lane will take on any document in the wrapper, a declaration, an interview summary, a decades-old amendment filed as a scan, and it takes a document fetch rather than a lookup. Reach for the first when it covers the question, and the second when it does not.
Before reading anything, there is a triage pass that says which office actions in a prosecution carry rejections under 101, 102, 103 or 112, which of them raise obviousness-type double patenting, which carry an eligibility flag, and which claims each rejection was directed to. On a long prosecution that is the difference between reading two documents and reading twenty.
Continuity comes back as a graph rather than a list. Parents, children, the relation on each edge in the Office's own vocabulary, continuation and continuation-in-part and divisional, foreign priority claims, and the earliest ancestors reached by walking the chain up rather than by trusting a single field. An application with no parents and no children says so, because empty is an answer here too.
Patent term adjustment arrives with the arithmetic shown: total days, the A, B and C delays, applicant delay, and the event history behind them. It stops there deliberately. It does not compute an expiration date, because expiration also turns on terminal disclaimers, maintenance fees and any extension under section 156, and a number that quietly ignores three of those is worse than no number.
Coverage
The structured rejection triage runs over office actions mailed from 1 October 2017. The office-action text reaches back considerably further, roughly to 2008 on measurement rather than by any published guarantee, so an empty triage result is not evidence that the text is unavailable.
Chapter V Patent Trial and Appeal Board
The Board runs three different kinds of proceeding and most tools cover one of them. This server covers all three: inter partes, post-grant, covered business method and derivation trials; ex parte appeals from examination; and interferences. A portfolio question that touches two of the three does not become two research projects.
Party matching respects sides. A search for a petitioner returns the proceedings that party filed, and a search for a patent owner returns the proceedings brought against it, so a company that is both never arrives as one undifferentiated pile. A list of up to two hundred trial numbers resolves in a single call, which is what a portfolio screen actually looks like.
The appeals lane carries the fields that make examination patterns visible: art unit, examiner, decision date, and the outcome as the Board recorded it. The docket comes back with real document categories, petition and preliminary response through institution decision, motions, oppositions, replies, the final written decision and any rehearing, filed by the Board, the petitioner or the patent owner, and the text of any of them can be read in the same conversation.
Search returns proceedings and status. It does not return which claims survived, and it declines to guess: that answer is in the decision, so the decision is what you are handed.
When the record is incomplete
Some dockets are sealed to the point where the Board's own paper is absent. The response says so rather than returning a bare empty result, because an empty list is otherwise indistinguishable from a proceeding in which no decision issued.
Chapter VI Final Petition Decisions
Petition decisions are where a prosecution's difficulty shows. A revival petition under 37 CFR 1.137 means the application went abandoned. A petition under 1.181 means somebody took a dispute with the examiner over the examiner's head. Several petitions on one application is a shape worth seeing before a diligence call rather than during one.
The dataset changed shape in the first week of October 2026. Until then the Office published about seven and a half thousand final decisions, nearly all of them denials. Read on 5 October 2026, it holds more than one point one million records: every petition decision since 2001, the automatic ePetition grants included. Roughly eight hundred and eighty thousand grants, two hundred and twenty thousand dismissals, seven and a half thousand denials, and under two thousand grants in part. A flag marks the final decisions, the old set, apart from the rest; the petition type codes now number 191 and can be listed; every record carries its document; and the decision text is searchable in full.
The server searches by applicant, application, patent, decision type, deciding office, art unit and date, with no query language to learn, and takes a final-only filter that narrows the million to the decisions somebody sat down and wrote. A first pass is a scannable index; the complete record, with every issue raised, the rules and statutes cited and the full timeline, is pulled only for the decision you actually want to read. The shape worth seeing is now read with the flag: a run of automatic grants on one application is routine, and a final decision among them is the one to open.
Three tools arrived with the data. FPD_get_petition_statistics returns counts and grant, dismissal and denial rates for any scope, by type, art unit, technology centre, rule, statute or dates, broken down by up to four dimensions, from the Office's own facet counts. FPD_get_dismissal_grounds reports where a scope's dismissals fall, with the newest dismissal decisions to read, because the Office records no reason field. FPD_list_petition_types gives the 191 petition type codes with the Office's own names and live counts, so "revival" or "PTA" becomes a code the other tools take.
Decisions are served verbatim. Where a decision was filed as an image it is transcribed rather than summarized, because legal research depends on the words as filed and a tidying pass would quietly change what you are quoting. Long decisions are paged rather than cut short, with the window snapping to page boundaries, so nothing is lost without being announced.
A genuine zero
When there is nothing to return, the answer says it is a genuine zero and not an error. It then says the one thing that keeps that claim honest: the server cannot tell a filter value that is merely absent from one that is misspelled, so it tells you to re-run without the narrowest filter to distinguish them.
Coverage is the second half of reading a zero. Petitions are held for applications and patents filed in 2001 or later, and since the October 2026 change the decisions run from 2001 as well, automatic grants included. So a zero on an application now says that no petition of any kind was decided on it, which is a stronger finding than the old set could make, and the server says which kind it looked for. The papers themselves are still in the file wrapper, two chapters back.
Dated capture, 5 October 2026, the week the dataset changed. The figures above are the Office's own facet totals that day, stated as floors and rounds, never as a served count.
Chapter VII USPTO Citations
The Office publishes the examiner's citations twice, through two APIs with two schemas, and neither contains the other. This server serves both, as two lanes with their own tools, and the working rule is to ask both: on one application the office action index returned 4 records and the enriched index 8, and on another the office action index returned 38 against 26. Which lane holds more is a property of the application in front of you.
The office action index is the citation list itself, transcribed from the examiner's Form 892 and the applicant's Form 1449. Its tools are Citations_search_oa_citations_minimal and Citations_search_oa_citations_balanced, and it alone carries the statutory basis a reference was cited under (legalSectionCode, 102 or 103 or 112), the type of action it appeared in (actionTypeCategory) and the broader inventory of what the applicant disclosed. It has no date field at all: the timestamp it carries is the date the record was loaded, never the date the action was mailed.
The enriched index is an AI-extracted reading of a subset of office actions. Its tools are Citations_search_citations_minimal and Citations_search_citations_balanced, with Citations_get_citation_details for one record in full and Citations_get_citation_statistics for aggregations, and it alone carries the relevance code (X, Y or A), the claims a reference was cited against, the passage locators the examiner pointed to, and officeActionDate, the one date either index has.
Ask one index for a field the other was built to hold and it does not return fewer results. It returns an error and none at all. The one field both answer to is the application number, which is the join.
Coverage, and what a zero means
The Office documents both APIs as covering office actions mailed from 1 October 2017. Both have been observed serving older prosecution: the reading below is on a 2012 application. So an older matter is worth querying rather than ruling out, and a zero in one lane is not evidence of no citations in the other. The examiner's name is in neither index; it is in the file wrapper, three chapters back.
US 2006/0075466 A1, on application 12849948. Both indexes hold it. Neither of them holds this much of it.
Neither half is the answer. The statutory basis without the claim mapping does not tell you what the reference was aimed at, and the claim mapping without the statutory basis does not tell you what it was aimed with. The pair does, and assembling the pair is two calls and a join on a number both sides carry.
Dated capture, 2 September 2026, re-read on the enriched side on 10 October 2026. Per-application readings on a named application, never a count of either index.
Colophon
The seven servers share the identifiers the Office already uses. A patent search ends by naming the calls that reach the live record; an application number joins the file wrapper to the petitions and to both citation indexes, a patent number joins it to the Board, and an art unit joins them all. The text server and the graph share a vocabulary of their own: one domain name is a pattern on the graph and a domain on the text server, and they answer it differently on purpose, one with the passage and one with its dates. So a single conversation can run from a search through the first office action and the petitions to a final written decision, and ask which rule governed at each step, without leaving it.
Connect all seven to Claude and put a real application through them.
Private beta. Access is by invitation: [email protected].
Private beta. Access is by invitation: [email protected].
Auto-renews at the then-current price until cancelled. Cancel any time from your account's billing portal; cancellation stops future renewals and does not itself refund the current period.
USPTO Full Bundle (Patent Search + PFW + PTAB + FPD + Citations + MPEP RAG + Hypergraph) is a research tool over public Office records. Retrieved document text is treated as quoted data and never as instruction, and an examiner's or a petitioner's characterization is an attributed position rather than an established fact. Summaries and rankings are AI-generated and require professional review; verify against the linked source document before relying on any output. Coverage floors differ per server and are stated in each chapter above, because a zero result means nothing until you know what the server can see.