Tollmint

Pay-per-call data and analysis for autonomous agents. Graded behavioural-science evidence, regulatory enforcement over deceptive design, dark-pattern detection, web fetch and geocoding. No account: quote and pay in the same HTTP round-trip over x402, or against prepaid credits.

Skills

• Dark Pattern Detector
  Scan a live page or raw HTML for deceptive design. Returns each detected pattern with the evidence quoted from the markup, the regulation it trips, and the honest alternative. Fully deterministic — no model call, so results are reproducible and the same page always scores the same. Returns a clean/review/fail verdict suitable for gating a deploy. Names the patterns that need a full user flow to assess rather than implying the site is clean.
  psychclass-Bgraded-evidence
• Cognitive Load Score
  Measure the mental effort a page demands before a decision can be made: form fields, competing choices, interactive density, and reading burden. Deterministic and reproducible — the same markup always scores the same, so the number is usable for tracking a page over time or against a competitor. Choice count is weighted logarithmically per Hick's law rather than linearly, which is the law's most common misapplication.
  psychclass-Bgraded-evidence
• Trust Signal Audit
  Check which trust signals a page carries and which it omits — pricing visibility, privacy and terms, contact routes, cancellation terms. Each gap is returned with the inference a visitor draws from its absence, because under ambiguity aversion missing information is read as adverse rather than neutral: a hidden price is inferred to be a high one. Deterministic, model-free, and reproducible.
  psychclass-Bgraded-evidence
• Behavioural Effects Corpus
  Query a curated corpus of behavioural science effects. Every entry carries its canonical citation, replication status (robust / mixed / contested / failed), effect size, and boundary conditions. Effects that failed replication are included and labelled, not hidden — filter with actionable_only to exclude them.
  psychclass-Bgraded-evidence
• Cognitive Bias Taxonomy
  Cognitive biases with their mechanism, the conditions that trigger them, documented countermeasures, and an exploitation-risk score. Every entry carries its citation and replication status. Use to audit a decision process, to harden a system against manipulation, or to check your own reasoning.
  psychclass-Bgraded-evidence
• Dark Pattern Taxonomy
  Deceptive-design taxonomy with the regulation each pattern trips (EU DSA Art. 25, UCPD, FTC ROSCA and the Negative Option Rule), the signals that detect it in markup or copy, a severity score, and the honest alternative that achieves the same business goal. Use before shipping conversion changes, or to audit a competitor.
  psychclass-Bgraded-evidence
• Deceptive Design Enforcement Record
  Regulatory enforcement actions over deceptive interface design, each citing the regulator’s own document: the law and article, the company, the penalty, and what was actually on the screen. Includes dismissals, vacated rules and findings of lawfulness, which mark where the line sits and which fine-only lists omit. Also reports patterns prohibited but never enforced, and cases excluded for want of a primary source. Use to tell whether a design choice has cost anyone money, and where.
  psychclass-Bgraded-evidence
• Interaction Design Laws
  Interaction-design laws with their formulae, sources, and design implications — Fitts, Hick, Jakob, Tesler, Doherty. Includes the common misapplications, since these are misquoted more often than they are wrong. Separate from the effects corpus because these are engineering constraints, not contested empirical claims.
  psychclass-Bgraded-evidence
• Score a Call to Action
  Score call-to-action text against a curated behavioural-science corpus. Checks verb framing (outcome versus effort), friction words introduced at the point of commitment, hedging, ownership framing, and length — each finding citing the effect it rests on and that effect's replication status, so a recommendation grounded in a contested result is visibly weaker than one grounded in a robust one. Separately flags urgency language the surrounding copy does not support, which is a regulatory exposure rather than merely a conversion cost. Deterministic: the same text scores identically every time, so two variants can be compared without sampling noise. Does not predict a conversion lift, because the corpus does not support that precision.
  psychclass-Bgraded-evidence
• Mechanical Style Score
  Measure how mechanically a piece of text reads across 19 stylistic signals: sentence-length variance (burstiness), em-dash rate, negative parallelism, copula avoidance, the characteristic AI vocabulary cluster, hedged superlatives, significance inflation, puffery, vague attribution, over-signposting, formatting tics, and more. Returns a per-sentence breakdown naming which signal each one tripped. Deterministic and model-free, so before-and-after comparisons across an edit are meaningful. NOT an AI detector — it reports stylistic properties, never authorship, and returns no verdict or probability. Authorship classifiers are unreliable and disproportionately misjudge non-native English writers.
  contentclass-B
• Template Convergence Detector
  Score how closely a page matches the documented 2025-2026 AI/template design fingerprint: the Tailwind indigo-violet gradient, a default sans with no display pairing, the badge/hero/three-card/steps/testimonial/pricing skeleton, glassmorphism, uniform border radii, emoji feature bullets, round-number social proof with no attribution, and abstract illustration in place of the product. Each finding quotes the markup that produced it and names the move that breaks the pattern. Deterministic and model-free. Reports template convergence, never authorship — the same fingerprint appears in human sites built from the same tutorials.
  designclass-B
• Fetch Page as Markdown
  Fetch a URL and return the main content as clean markdown, with navigation, cookie banners, footers, and related-post chrome removed. Extracts title, byline, publication date, and excerpt. Content detection is driven by link density rather than class names, so it survives CSS refactors. Reports how much of the page was retained and whether it had to fall back to the full body, so you can tell a clean extraction from a guess. Honours robots.txt; refuses private addresses and pages behind authentication.
  webclass-C
• Scan Copy Against Regulatory Rules
  Check marketing copy, or a live page, against 225 curated regulatory rules covering AI-content disclosure, advertising claim substantiation, subscriptions and price display, unsolicited messaging, accessibility, and automated data collection. Every finding cites the instrument and article plus a primary-source URL, and reports whether that URL resolved when the corpus was built. Crucially it also returns what is NOT law: rules commonly believed to be in force that were vacated, enjoined, or superseded — the FTC Click-to-Cancel rule, the California election-deepfake statutes, the DOT ancillary-fee rule. A scanner built by prompting a model asserts all three as live and sends you to rewrite compliant copy. Dead rules never enter the analysis and are reported separately with the reason they died. Reports rule conflicts and citations; does not determine liability.
  verifyclass-Agraded-evidence
• Geocode
  Convert a free-form address or place name into coordinates with normalised address components, a calibrated confidence score, and precision classification (rooftop / street / locality / region / country). Deduplicates OSM elements and flags ambiguous queries with the competing alternatives. Data from OpenStreetMap.
  refclass-C
• AI Disclosure Obligations
  Binding obligations worldwide to disclose, label, or watermark AI-generated content — EU AI Act Art. 50, California SB 942 and AB 2013, China’s 2025 labelling Measures, South Korea’s AI Basic Act, India’s 2026 IT Rules. Each row gives the exact trigger condition, what must be done, the exemptions, the penalty ceiling, and whether anyone has been enforced against yet. Includes forward dates for obligations already enacted but not yet biting, and records the California election-deepfake laws that were struck down in 2025 and are still widely cited as live. Query before shipping generated content into a jurisdiction.
  legalclass-Bgraded-evidence
• Marketing Claim Substantiation
  Which marketing claims legally require evidence, and what evidence suffices. Covers superlatives, health and efficacy, "free", environmental and carbon-neutral claims, reference and "was/now" pricing, "up to X%", reviews and testimonials, country of origin, guarantees, and AI-capability claims. Each row gives the regulator’s own evidentiary standard, the common way advertisers fail it, and enforcement with penalties where it exists. Also records five widely repeated "you can’t say that" rules that no instrument actually supports.
  legalclass-Bgraded-evidence
• Subscription and Price Display Law
  Rules on subscription sign-up, auto-renewal, cancellation, and total-price display across US federal, US states, EU, UK, France and Germany — with the numbers: how many days before renewal a reminder is due, how many clicks cancellation may take, what must be in the price. Critically, it records that the FTC’s "Click-to-Cancel" rule was vacated on 2025-07-08 before it ever took effect, along with the vacated DOT ancillary-fee rule and the twice-delayed UK subscription regime. A model working from training data will assert all three as live law.
  legalclass-Bgraded-evidence
• Outreach and Messaging Compliance
  When a commercial email, SMS, or call may lawfully be sent, by channel and jurisdiction. The field that matters is consent_model: the US is opt-out for email, Canada and the EU are opt-in, and Germany applies opt-in to B2B while France and the Netherlands do not. Applying the US model abroad is a breach. Includes opt-out deadlines, permitted calling hours, whether a private right of action exists (the difference between a regulator fine and a class action), and the FCC one-to-one consent rule vacated a day before it took effect.
  legalclass-Bgraded-evidence
• Web Accessibility Obligations and Enforcement
  Accessibility duties and the cases brought under them: ADA Titles II and III, Section 508, the European Accessibility Act in force since 2025-06-28, EN 301 549, UK and Canadian regimes — with the WCAG version and conformance level each one actually incorporates, the microenterprise and undue-burden carve-outs, and compliance dates still ahead. Enforcement rows include the FTC’s $1M action against an accessibility overlay vendor and a case where a site running an overlay was sued anyway, plus defence wins that mark the limits.
  legalclass-Bgraded-evidence
• Automated Data Collection Legality
  What courts have actually held about scraping, crawling, and training-data acquisition. Answers questions in the form an engineer asks them: does ignoring robots.txt create liability, does a browsewrap bind a bot, is public data outside the CFAA. Records binding scope honestly — a Ninth Circuit holding is not US law — and tracks the full history of hiQ v LinkedIn, which is routinely cited as making scraping legal and which hiQ ultimately lost on contract. Covers the EU text-and-data-mining exceptions and their opt-out.
  legalclass-Bgraded-evidence
• Conversion Psychology Page Audit
  Audit a live page or raw HTML for conversion psychology. Returns measured cognitive load, detected dark patterns with the specific regulation each trips and the evidence found in the markup, a trust-signal gap list, and ranked findings tied to graded behavioural effects. Findings may only cite effects from our corpus, and citations are attached from our data rather than generated — so a reference is never invented. Names the patterns a static page cannot assess rather than implying the site is clean.
  psychclass-Agraded-evidence
• Humanize Text
  Rewrite text for tone, cadence, and brand voice while preserving meaning exactly. Returns a before/after analysis: burstiness, hedge and transition density, abstract-noun ratio, plus the specific sentences that read mechanically and which signals they tripped. Reports signals the rewrite failed to clear rather than hiding them. Declines academic manuscripts and requests to defeat AI detection — enforced by structural checks on the submitted text, not by a policy the caller self-certifies against.
  contentclass-A

How to call

https://api.tollmint.com

Listed in

Directories this entry was found in.