The agent for IP attorneys that learns your way.

Cyrus is the agentic IP-law workspace. Citation-grounded. Plan-driven. Built for the way your firm practices.

Cyrus

I'm working on drafting an office action response for 18/342,109. Read the prosecution history, pull the latest office action, review and analyze the cited arts, and tell me the top 3 strongest deficiencies in the prior art cited by the examiner that we can argue — or amend the claims around — to overcome the current rejections.

4 tools · 6.4s· showing all steps

Listing files for matter 18/342,109

18 papers on file · Non-Final OA mailed 2025-03-12 · IDS · 2 cited references attached

1.42s

Loading file Non-Final OA mailed 2025-03-12

§103 over US 10,452,978 (Vaswani) i/v US 11,042,796 (Lewis) · claims 1–14 · 9 limitations cited

1.31s

Building claim dependency graph for matter 18/342,109

9 elements · 13 dependent claims · element (c) flagged as the input-conditioned gating step

0.95s

Asking wiki about matter 18/342,109

Comparing Vaswani col. 7–9 + Lewis claims 1–4 to each claim 1 element · 3 disclosure gaps identified

2.72s

I pulled the file wrapper for 18/342,109, parsed the Non-Final OA mailed 2025-03-12, and ran an element-by-element comparison of claim 1 against the cited art.

Top 3 deficiencies in the Examiner's §103 over US 10,452,978 (Vaswani) in view of US 11,042,796 (Lewis):

1. The conditional gating step is not disclosed. Vaswani col. 8:14–23 1 shows only unconditional self-attention; Lewis claim 2 2 covers post-retrieval rerank. Neither reference teaches the input-conditioned weighting gate recited in claim 1, element (c) — the Examiner's combination assumes a feature absent from both refs.

2. Lewis is structurally incompatible with Vaswani's training loop. Lewis 2 operates on already-retrieved candidates at inference and produces gradient-blocked output; the application's pipeline runs at training time. The OA never explains how Lewis's output would back-propagate into Vaswani's encoder.

3. The combination teaches away from itself. Neither reference describes the engineering bridge the Examiner relies on — the gradient-flow conflict between Lewis's frozen rerank and Vaswani's trainable encoder means a person reading both would not naturally combine them.

Amend-around path. Narrow element (c) to “gating weight derived from a learned function of the input context” — supported by spec ¶¶ 0042–0048 3, no §132 new-matter risk.

Proposed next step · Draft the §1.111 response

1Structure rebuttal around the missing gating-step disclosure (deficiency #1)

2Distinguish Lewis on pipeline-stage + gradient-flow conflict (deficiency #2 + #3)

3Add narrowing amendment to element (c) anchored to spec ¶¶ 0042–0048

4Draft in firm voice from matter 2024-PA-0084 (won at first response, same Art Unit)

Office action response · 18/342,109×

Ask a follow-up, or paste an office action…

Cyrus 1.0

Cyrus

Drafting an OA response for 18/342,109. Read the file wrapper, pull the latest OA, analyze the cited art, and give me the top 3 deficiencies we can argue or amend around.

4 tools · 6.4s

Listing files for matter 18/342,109

18 papers · OA + IDS

1.4s

Loading Non-Final OA

§103 Vaswani i/v Lewis

1.3s

Building claim dependency graph

9 elements · 13 deps

0.9s

Asking wiki about matter

3 disclosure gaps

2.7s

Top 3 deficiencies — §103 over Vaswani i/v Lewis:

1. The conditional gating step in claim 1 element (c) is not disclosed — Vaswani col. 8:14–231 shows only unconditional self-attention; Lewis2 is post-retrieval rerank.

2. Lewis's gradient-blocked rerank can't back-propagate into Vaswani's encoder — pipeline-stage mismatch the OA never addresses.

3. Combination teaches away from itself — no engineering bridge between the two refs.

Next step · Draft §1.111 response

1Structure rebuttal around the missing gating step

2Add narrowing amendment to (c) · spec ¶¶ 0042–00483

3Draft in firm voice from matter 2024-PA-0084

OA response · 18/342,109×

Ask a follow-up…

Cyrus 1.0

The plan feature

Plan-first. Every step traceable.

Cyrus reads your prompt, picks the right tools, and shows you the plan before it runs. Every action is logged, every citation sourced.

PTAB research · skilled-artisan recognized-benefit doctrine

5 PTAB Final Written Decisions where a skilled artisan would recognize a benefit of a proposed prior art modification even though the modification may impede a function of the primary reference, and the claim is found unpatentable as obvious. 2025–2026, preferably 2026.

Plan3 of 5 done

Searching IPR (PTAB) for skilled-artisan recognized-benefit doctrine

Filtered to FWDs in 2025–2026 with claims found unpatentable as obvious · 247 candidates

Verifying 247 IPR candidates against query

Narrowed to 18 FWDs where Patent Owner argued the modification would impede a function

Reading IPR documents for Board's benefit-recognition framing

Verified passages anchored to FWD page numbers · 5 cases survive

Getting PDFs for the 5 IPR cases and highlighting cited passages

Yellow-highlight every cited passage · 3 of 5 PDFs annotated

Compile report with anchored citations + Highlighted-FWDs folder

Found · 5 FWDs (2025–2026)

Thermaltake v. Chen — IPR2024-01230 · ’336 patent · claims 1–5 unpatentable2026-02-17

Biofrontera v. Sun Pharma — IPR2024-01312 · ’028 patent · all challenged claims unpatentable2026-02-23

Imperative Care v. INARI Medical — IPR2024-01157 · ’011 patent · all challenged claims unpatentable2026-01-16

Solaris Oilfield v. Masaba — IPR2024-01179 · ’689 patent · claims 1–25 cancelled2026-01-26

FOX Factory v. SRAM — IPR2024-00492 · ’207 patent · claims 1–3, 7, 8, 10–12 unpatentable2025-08-04

One workspace. Seven products.

Each grounded in a different sources — the MPEP, patents and publications, every PTAB case, every CAFC orders & opinions — and all wired into Cyrus.

Cyrus · agentic IP partner

Plan-first agent for patent prosecution

Citation-grounded. Plan-driven. Agentic from day one — searches, drafts, charts, and opinions, all from one workspace.

Final OA on 18/342,109 just dropped — §103 rejection over Lee + Tanaka + Chen, claims 1–5. I have seven days. Read all three refs, chart each combination, find which arguments survived in my prior §103 responses, and draft an amendment + response in my voice.

●5 tools · file wrapper · cited refs parser · argument archive · claim chart · amendment gen · 9.4s

Response strategy — 18/342,109 (matter 2024-PROS-0142):
1. §103 Lee + Tanaka — Tanaka’s batched training does not teach claim 1’s incremental schedule; argue hindsight per1 Plantronics v Aliph (CAFC 2014).
2. §103 Lee + Chen — Chen teaches away from sub-batch normalization (col. 7:22–48); KSR teaching-away applies2.
3. Argument precedent — three of your prior wins (3 16/834,221, 17/123,456, 17/889,012) paired the same hindsight + teaching-away framing successfully.
4. Amendment — narrow claim 1 to incremental schedule with sub-batch normalization — clearly outside Lee + Chen’s disclosure. Cost: ~6 attorney hours.

Proposed next step · Draft amendment + response in firm voice

1Re-validate Tanaka batched-training read vs spec ¶¶ 41–53

2Pull Plantronics + KSR teaching-away quotes for arg block

3Cite 3 prior wins (16/834,221, 17/123,456, 17/889,012)

4Draft amendment + arguments in firm voice · 2024-PROS-0142

OutputsDOCXOA-response-18-342-109.docxPPTXclaim-chart-Lee-Tanaka-Chen.pptxPDFargument-precedent-cite-list.pdf

Vault · privileged matter store

Your matters, your privilege boundary

Centralized intake, audit-logged access, privilege-by-default.

2025-PA-0142

Arclight Robotics Inc.

18/901,553 · 14 files

Privileged

2024-PA-0084

Halogen Optics LLC

17/998,221 · 27 files

Privileged

2024-PA-0079

Helix Biosystems Corp.

17/882,406 · 41 files

Privileged

2025-LT-0031

Sentinel Imaging Inc.

N.D. Cal. 25-cv-00041 · 89 files

Privileged

2025-PA-0166

Northwind Energy LLC

18/944,012 · 9 files

Privileged

Patents · US grants and publications

Prior art search that reads like a partner did it

Every US patent. Every publication. Every cited paper.

US10452978B2Claim 1

Google LLC · Vaswani

0.94

US11138392B2Abstract

Google · Macherey

0.89

US10789529B2Claim 4

Microsoft · Zhang

0.83

US11232358B1Spec ¶42

Amazon · Kolesnikov

0.78

US11042796B2Claim 2

Meta · Lewis

0.71

US10963493B1Claim 1

Salesforce · Socher

0.62

MPEP · canonical patent doctrine

The MPEP at your fingertips — verified, current

Patent procedure, citation-grounded.

When is a claim limitation interpreted as means-plus-function under § 112(f), and how do I avoid invoking it inadvertently?

Apply the three-prong test from MPEP § 2181.I (post-Williamson):

Nonce term. The claim uses "means" or a non-structural placeholder — "module," "mechanism," "element," "unit," "device for."
Functional language. The term is modified by functional language, typically introduced by "for" or "configured to."
Insufficient structure. The claim itself does not recite sufficient structure or material to perform the claimed function.

If all three are met, § 112(f) is invoked and the claim is construed to cover only the corresponding structure disclosed in the specification plus equivalents (MPEP § 2181.II). To avoid invocation, use structural terms — "processor," "circuitry," "memory storing instructions to" — rather than functional placeholders.

Williamson v. Citrix Online, LLC, 792 F.3d 1339 (Fed. Cir. 2015) (en banc) CAFC overruled the strong presumption that absence of "means" avoids § 112(f).

IPR · post-grant agent

Win institution before you file

PTAB-grade petition and response work.

#194%All UnpatentableComp. Arch.

IPR2024-00345

Acme Networks v. Quantum Labs

#287%All SurviveNetworking

IPR2024-00891

Sentinel Imaging v. Helix Bio

#381%Mixed FWDBiotech

IPR2023-01124

Northwind Energy v. Arclight Robotics

CAFC · Federal Circuit precedent

Find the case that decides yours

18k+ Federal Circuit opinions, indexed.

Phillips v. AWH Corp.

415 F.3d 1303 (Fed. Cir. 2005)

Claim construction · cited ~4,200 times

Markman v. Westview

52 F.3d 967 (Fed. Cir. 1995)

Construction is a question of law · ~8,400 cites

Vitronics Corp. v. Conceptronic

90 F.3d 1576 (Fed. Cir. 1996)

Intrinsic-evidence hierarchy · ~3,100 cites

Innova/Pure Water v. Safari

381 F.3d 1111 (Fed. Cir. 2004)

POSITA-at-filing rule · ~1,800 cites

Texas Digital v. Telegenix

308 F.3d 1193 (Fed. Cir. 2002)

Dictionaries as fallback (limited by Phillips)

Notes · semantic search over your archive

Find what you've written before

Every claim, plan, and brief you've drafted, one query away.

antecedent basis fix for claim term first introduced in a dep claim3 hits

§112(b) antecedent-basis fixplan

… reframe “the said modulator” as “the modulator of claim 1” when the term is first introduced in a dependent claim to cure antecedent-basis indefiniteness …

edited 9 days agoMPEP § 2173Williamson v. Citrix

Response to OA — claim 14 indefinitenessbrief

… Applicant respectfully traverses the rejection. The term “control signal” finds antecedent basis in dependent claim 14, which depends from independent claim 1 reciting a signal generator …

edited 6 wk agoMatter 2024-PA-0079

Claim 1 — adaptive controller (rev. 3)claim

… 14. The controller of claim 1, further comprising a feedback path coupled to the modulator, the feedback path configured to adjust a control signal based on …

edited 4 mo agoApp 17/998,341

Cyrus · capabilities

Six jobs Cyrus does end-to-end.

Each grounded in a different combination of the platform's tools. Each verified, sourced, and signable.

Searches prior art exhaustively

Every US patent and publication, foreign equivalents, and peer-reviewed scientific literature — ranked by §102 anticipation and §103 obviousness relevance.

dynamic clause weighting attention transformer

US 10,452,978 · Google

Attention-based sequence transduction neural networks

US 11,138,392 · Google

Machine learning models for translation prediction

US 10,789,529 · MS

Neural info retrieval with dynamic re-weighting

US 11,232,358 · Amazon

Graph-attention retrieval over citation networks

US 11,042,796 · Meta

Multi-stage reranking using cross-encoder scores

US 10,963,493 · Salesforce

Adaptive passage scoring with session-level prior

Drafts patent applications

Reads invention disclosures, drafts the spec, generates dependent claim trees — every limitation traceable.

Claim 1 · Independent

A retrieval system comprising an encoder neural network configured to receive an input sequence and a dynamic clause-weighting gate…

Claim 2 · Dependent

The system of claim 1, wherein the encoder neural network includes a transformer block configured to apply multi-head self-attention.

Claim 3 · Dependent

The system of claim 1, wherein the dynamic clause-weighting gate selects salient sub-sequences via a learned scoring function.

Responds to Office Actions

§101 / 102 / 103 / 112 rejections with cited responses and claim amendments.

Examiner · §103

Claim 1 obvious over Vaswani in view of Lewis…

Cyrus · TSM rebuttal

Vaswani discloses unconditional self-attention; lacks the dynamic clause-weighting gate. Per MPEP §2143.01…

Amendment · claim 1

Add “wherein the gate weighting is conditioned on a learned query embedding” to traverse §103 obviousness.

Auto-judge claim elements against the corpus

Decompose claim 1 into its elements, run top-K retrieval per element, then judge each candidate teaches / lacks / partial — with shimmer while judging and reject-flash when an element fails.

Claim 1 element

US 10,452,978 (Vaswani)

US 11,042,796 (Meta)

US 10,789,529 (MS)

encoder neural network with stacked self-attention

Teaches¶0034

Partial¶0019

Lacks

decoder masked-attention sublayer

Teaches¶0041

Lacks

cross-attention between encoder + decoder

Teaches¶0047

Partialclaim 3

Lacks

positional encoding of input sequence

Teaches¶0052

Teaches¶0018

Partial

dynamic clause-weighting gate at retrieval time

Judging…

Partialclaim 2

Lacks

Runs FTO clearances

Maps the landscape around your product, identifies risk, surfaces design-arounds.

US 11,042,796Risk

US 10,789,529Watch

US 11,138,392Clear

EP 3,887,221 A1Clear

US 11,232,358Watch

US 10,963,493Clear

JP 7-298-441Risk

US 10,891,352Clear

2 risks12 patents reviewed · 3 design-arounds drafted

Auto-track every matter on your docket

Watches every prosecution + PTAB proceeding. Pulls + renames docs automatically. Push + email reminders the moment something lands.

Docket activity12 watched

18/342,109Non-Final Office Action mailed2 hrs agopulled + renamed

IPR2024-00345Patent Owner Response due in 7 daysdeadlinereminder sent

IPR2024-00210Petitioner Reply filed14 min agosynced to vault

17/998,341IDS deadline in 3 daysdeadlinereminder sent

NotificationsPushEmail

The prior art module

Every US patent, and publication.

Grants and application by claims, abstract, detailed description, background, CPC. Search in natural language, filter by date and technology area, and every response is grounded in the exact passage of the cited document.

Open the prior art module →

app.mpepai.com/prior-art-search

Prior art search

US patents + applications + scientific papers

Claim 1 · decomposition

Top K3 Auto-judge running

Element

REF 1

REF 2

REF 3

[1a]an encoder neural network configured to receive an input sequence comprising a plurality of tokens

Query: encoder neural network receive input sequence of tokens

Strong coverageAuto-judge: complete 3 / 3

TeachesPDF

US10452978B2

an encoder neural network configured to receive the input sequence and to process the input sequence to generate a respective encoded representation

TeachesPDF

US11138392B2

wherein the encoder neural network comprises a stack of self-attention layers processing the token sequence

PartialPDF

US10726784B2

the encoder receives a sequence of input embeddings, each embedding corresponding to a token in the input

[1b]a self-attention sub-layer applying multi-head attention over the encoded representations

Query: multi-head self-attention over encoded representations

Strong coverageAuto-judge: running 1 / 3

TeachesPDF

US10452978B2

applying a self-attention mechanism over the input sequence to compute, for each position, a weighted combination of values derived from the entire sequence

US10719587B2

[Claim 5]

the attention layer comprises multiple parallel attention heads, each computing scaled dot-product attention over the input...

0.87

US20210073644A1

[Spec ¶ 0042]

in some embodiments, a plurality of attention heads operate in parallel on the encoded sequence...

0.76

[1c]a learned weighting gate conditioned on the input sequence and applied prior to the softmax normalization

Query: learned weighting gate input-conditioned before softmax

Moderate coverageAuto-judge: 1 / 3 validated

PartialPDF

US10789529B2

a learned weighting gate adjusts retrieval scores conditioned on the input query context, applied before softmax normalization

US11042796B2

[Claim 7]

a re-ranking component configured to re-score the retrieved passages using a cross-encoder before generation

0.71

US20210073644A1

[Spec ¶ 0042]

the gating function comprises a learned linear projection of the input context, followed by a sigmoid non-linearity...

0.78

[1d]a decoder neural network applying cross-attention over the encoder outputs to autoregressively generate an output sequence

Query: decoder cross-attention autoregressive output sequence

Strong coverageAuto-judge: running 2 / 3

TeachesPDF

US11138392B2

a decoder neural network applying multi-head attention with a learned per-head temperature parameter over the encoder outputs

PartialPDF

US10452978B2

the decoder autoregressively generates each output position conditioned on previously generated outputs and the encoder representations

US10726784B2

[Spec ¶ 0089]

the decoder layer attends to encoder outputs via a cross-attention sub-layer interleaved with masked self-attention...

0.79

[1e]wherein the system generates the output sequence based on the weighted combinations from the self-attention sub-layer

Query: generate output sequence from weighted attention combinations

Moderate coverageAuto-judge: 1 / 3 validated

US20220198250A1

[Spec ¶ 0089]

the disclosed graph-attention retrieval engine traverses a citation network to score candidate prior-art documents...

0.66

PartialPDF

US10452978B2

producing an output sequence based on the weighted combinations from the self-attention layer applied to the encoded representations

US11042796B2

[Claim 1]

generating a response using a retrieval-augmented generator conditioned on retrieved passages from a corpus

0.69

ValidatedJudgingRejectedPreview

app.mpepai.com/ipr

IPR · PTAB search

skilled artisan recognized benefit + modification impedes function⌘ K

FiltersOutcome · All Unpatentable, Mixed FWD ×Grounds · § 103 ×FWD date · 2025–2026 ×

5 cases · sorted by relevance Sort: Relevance ↓

#194%

IPR2024-01230

U.S. Pat. 10,690,336

Thermaltake Technology v. Chen, Chien-Hao

Filed Sep 2024 · FWD Feb 17, 2026

§ 103 · Echazarreta i/v Lai/Hasegawa

Claims 1–5 unpatentable as obvious

All UnpatentableSemiconductorsIPR52 docs

#291%

IPR2024-01312

U.S. Pat. 11,697,028

Biofrontera Inc. v. Sun Pharma Indus.

Filed Oct 2024 · FWD Feb 23, 2026

§ 103 · Lundahl i/v Larsen

All challenged claims unpatentable

All UnpatentableBiotechIPR47 docs

#388%

IPR2024-01157

U.S. Pat. 11,697,011

Imperative Care v. INARI Medical

Filed Jul 2024 · FWD Jan 16, 2026

§ 103 · Schaffer + secondary refs

All challenged claims unpatentable

All UnpatentableMedical DeviceIPR39 docs

#485%

IPR2024-01179

U.S. Pat. 11,780,689

Solaris Oilfield v. Masaba

Filed Aug 2024 · FWD Jan 26, 2026

§ 103 · cites Corephotonics tradeoff doctrine

Claims 1–25 cancelled

All UnpatentableMech. Eng.IPR36 docs

#582%

IPR2024-00492

U.S. Pat. 7,147,207

FOX Factory v. SRAM

Filed Feb 2024 · FWD Aug 4, 2025

§ 103 · Yi i/v JP’279

Claims 1–3, 7, 8, 10–12 unpatentable

Mixed FWDMech. Eng.IPR44 docs

The IPR module

Every PTAB case, searchable.

Every Final Written Decisions, citation-grounded. Filter by tech area, outcome, statutory grounds, and APJ panel — every holding traceable to the exact passage in the FWD PDF.

Open the IPR module →

Drop in, draft out

Speaks the file formats your firm files in.

Drop an office action PDF, invention disclosure, or PTAB petition. Cyrus reads it, decomposes the claims, and ships back Word, PowerPoint, and PDF — formatted for filing.

See supported formats →

Cyrus · intake

Reading office actionextracting

PDF

OA-18-342-109.pdf

14 pages · USPTO non-final · received 2026-05-18

82%

Detected §103 rejection over Vaswani in view of Lewis

Decomposed claim 1 into 14 elements

Cross-referencing claim mapping table

Ready to file · 3 outputs

Response package

App 18/342,109 · drafted in firm voice from matter 2024-PA-0084

DOCX

response-§103.docx

7 pages · Graham + KSR analysis · firm signature block

PPTX

claim-chart-1.pptx

14 elements × 2 references · element-by-element grid

PDF

IDS-PTO-SB-08.pdf

PTO/SB/08 · 14 references · ready for EFS-Web

Cyrus · intakeextracting

PDF

OA-18-342-109.pdf

14 pages · non-final OA

82%

Detected §103 over Vaswani i/v Lewis
Decomposed claim 1 into 14 elements
Cross-referencing claim mapping table

Ready to file · 3 outputs

Response package

App 18/342,109 · firm voice from matter 2024-PA-0084

DOCX

response-§103.docx

7 pages · Graham + KSR

PPTX

claim-chart-1.pptx

14 elements × 2 refs

PDF

IDS-PTO-SB-08.pdf

PTO/SB/08 · EFS-Web ready

US 11,887,367 — Claims.docx

Specification

Claims

Listing of Claims

U.S. Patent No. 11,887,367 B1 · Using Machine Learning to Train and Use a Model to Perform Automatic Interface Actions Based on Video and Input Datasets · Baker et al., OpenAI Opco LLC · Reissue draft under 37 C.F.R. § 1.173

1.(Original)

A method for training a machine learning model to perform automated actions, comprising:receiving unlabeled digital video data;generating pseudo-labels for the unlabeled digital video data, the generating comprising: receiving labeled digital video data; training a first machine learning model including an inverse dynamics model (IDM) using the labeled digital video data; and generating at least one pseudo-label for the unlabeled digital video data, wherein the at least one pseudo-label is based on a prediction, generated by the IDM, of one or more actions that mimic at least one timestep of the unlabeled digital video data, and the prediction of the one or more actions is generated based on a non-causal combination of past information and future information within the unlabeled digital video data, the past and future information being relative to one or more reference frames within the unlabeled digital video data;adding the at least one pseudo-label to the unlabeled digital video data to form pseudo-labeled digital video data; andfurther training the first machine learning model or a second machine learning model using the pseudo-labeled digital video data to generate at least one additional pseudo-label for the unlabeled digital video.

2.(Currently amended)

The method of claim 1, wherein the IDM or machine learning model is trained to generate one or more predicted actions to be performed via a graphical user-interface overlay rendered atop a live application windowTab ↹without invoking an operating-system input event on the host machine.wherein the overlay is generated by a second neural network distinct from the IDM.and refreshed at a rate exceeding sixty frames per second.

Style preservation

Drafting that reads like you wrote it.

Cyrus learns your firm's signature phrasings, claim structure, and citation conventions. Press Tab to accept — the next phrase appears before you finish typing.

Built for the way IP firms actually work.

Search the data that wins your case.

Every canonical corpus pre-indexed with frontier embedding models — exact-passage semantic match across MPEP, PTAB, CAFC, the full US patent corpus, and peer-reviewed papers. Competitors fall back to web search; we mathematically reach the closest passage every time.

Explore the corpora →

Search every embedded corpus at once

Corpus 5 active

Privileged. Audit-logged. By default.

Attorney-client and work-product privilege baked in. Every access logged, every download attributed. Compliance-ready for state-bar audits and SOC 2 (in progress).

Security and privilege →

Found14 access events

By3 attorneys, 1 paralegal

Privilegeboundary intact

Citation-grounded by construction

Every output Cyrus produces — every paragraph in an opinion letter, every cell in a claim chart, every line in a §103 response — carries an inline N badge linking to the source. Nothing ungrounded ships.

How citations work →

Under MPEP §21431 the examiner must articulate a reason with rational underpinning. Therasense2 holds disclosed-but-not-enabled is not anticipatory.

1MPEP § 2143 · Chapter 2100pp. 12

“The examiner must articulate a reason with rational underpinning to support the legal conclusion of obviousness.”

Roadmap

ShippedMay 17, 2026

Per-stage activity trace + findings UI

ShippedMay 16, 2026

Claim decomposition + auto-judge matrix

ShippedMay 12, 2026

Sub-second retrieval across legal corpora

ShippedMay 6, 2026

Prior-art search across US patents + applications

In progressNow

Privileged matter vault

In progressNow

Custom agent skills

PlannedQ3 2026

Public API for agent + corpus access

See the full roadmap →

The agent for IP attorneys that learns your way.

Free Beta. Join the waitlist or try MPEP and CAFC search free — no sign-up required.

Join the waitlist →Try MPEP free →