What's New: ChatGPT Apps & Ranking Changes
Every new ChatGPT app we add, every retest that moves a score, and every status change — dated, slug-anchored, and machine-readable. This page is the canonical changelog for the directory and the weekly feed of new apps inside ChatGPT.
New ChatGPT apps this week
The most recently tested apps in the directory — the closest thing to a live "new ChatGPT apps" feed.
- Daft.ieSearch Irish property listings conversationally inside ChatGPT.Last reviewed Aug 24, 2026
- BayutSearch UAE property listings by describing what you want, inside a ChatGPT conversation.Last reviewed Aug 24, 2026
- LucidFind, summarise, and generate Lucid diagrams and boards from a ChatGPT conversation.Doc-checked Aug 24, 2026
- WeightWatchersGLP-1-aware meal planning and food guidance inside ChatGPT Health, for WeightWatchers members.Last reviewed Aug 24, 2026
- Function HealthBring your own lab results into a ChatGPT conversation so health answers are grounded in your biology rather than population averages.Doc-checked Aug 24, 2026
- SciteGround answers in peer-reviewed literature, and see whether a finding is supported or contradicted by later research.Doc-checked Aug 24, 2026
Editorial changelog
Newest first. Each entry is dated and, where applicable, anchored to the specific apps that changed.
Corrected the site's plugin terminology and opened a surface for it. OpenAI migrated the App directory to the Plugin directory in July 2026, and "plugin" now names the installable package — which can contain skills, an MCP server, or both — while "app" still names the MCP-backed integration inside it. Our existing plugin content said the opposite (that plugins were retired and replaced by apps), which was accurate until mid-2026 and is not any more; the glossary entry and the apps-vs-GPTs FAQ have been rewritten against OpenAI's own developer documentation. Added /chatgpt-apps-vs-plugins, a layer-by-layer disambiguation page with a vocabulary translator and five primary-source citations, and /best-chatgpt-plugins, a ranking surface for the plugin head term that leads with the terminology answer and hands off to the full table. Added two glossary entries (plugin directory, skills). Stated the scope boundary explicitly: we rank MCP-backed plugins, not skills-only plugins, because our seven criteria have no meaning without a service behind the listing. Extended llms.txt with a terminology-mapping section so agents stop conflating the layers. No slugs changed.
Added two derived surfaces. /alternatives/<slug> gives every app a "what should I use instead" page built from its curated alternatives plus the highest-scoring apps in its categories; where nothing outscores the app, the page says there is no upgrade rather than promoting a weaker substitute. /chatgpt-app-availability collects documented region limits and plan requirements — 13 apps with a regional restriction, 6 with a plan requirement — and states plainly that the remaining apps have no documented constraint, which is not the same as confirmed global availability. There is no official per-country availability matrix for ChatGPT apps, so we do not publish one.
Wave 1 build-out. Added Yelp, Resy, and Thumbtack with primary-source evidence (catalog 65 to 68); rejected Supabase and Vercel, which surfaced only as "Sign in with ChatGPT" partners rather than apps running inside a conversation. Shipped /llms-full.txt, a single-file export of the entire dataset for agents that will not crawl page by page. Expanded direct-answer FAQ pages from 34 to 60 intents, each grounded in the live editorial score, and rewrote the restaurant-reservations answer that the new apps had made incomplete. Added /chatgpt-plugin-directory-tracker: a sourced timeline of the official directory with every count labeled Verified count or Reported estimate, because OpenAI publishes no running count and the widely-quoted ~300 figure is press reporting rather than an official statistic.
Ingestion round two, catalog 68 to 72. Added Klarna (cross-merchant product search on Klarna's own MCP server, 13 markets), Ticketmaster and SeatGeek (live-event discovery and ticketing), and Uber Eats (delivery, separate from Uber's ride-hailing app). Also published two documented negatives: there is no native Strava app for ChatGPT (its official connector shipped Claude-first and Strava's API policy restricts AI use of its data) and no native Xero app (its official connector is in Claude's directory, and its own AI work went to an in-product agent). Both are heavily advertised as integrations by third-party bridge services, so saying plainly that nothing native exists is more useful than a shortlist. Rejected as out of scope: Zendesk, Intercom, QuickBooks, and Xero reachable only via self-hosted or third-party MCP hosting; Walmart, whose ChatGPT presence is a commerce and payments tie-in rather than a directory app; and Headspace, Calm, and Whoop, for which no announcement could be found.
Ingestion round three, catalog 72 to 74. Added Redfin (conversational home search with neighborhood and market context, first-party announcement 2026-02-06) and Realtor.com (deliberately narrower — pre-search affordability and orientation, then routes buyers back to its own site, which is reflected in its lowest sub-score). Zillow now cross-links to both. Also published the catalog's most common boundary case as its own answer: there is no SEO app in the plugin directory. Semrush runs an official MCP server ChatGPT can connect to, but a vendor-hosted MCP server you wire up yourself is infrastructure, not a published listing — no directory entry, no review, setup required. Getting that distinction right matters more as more vendors ship MCP servers without shipping plugins.
Ingestion round four, catalog 74 to 77 — the hotel-chain cluster. Added ALL Accor (launched 2026-01-29, reported 20+ languages), IHG (2026-06-03, discovery then handoff to IHG's own booking flow), and Radisson, which carries the thinnest documentation of any entry in the catalog: one trade report, no first-party announcement, no precise launch date. Its review states that plainly and explains that score differences among the three hotel apps reflect documentation quality rather than a real quality gap. All three are single-brand by design and cannot compare a property against other groups. Deliberately NOT added: Hilton, which announced a ChatGPT app on its Q1 2026 earnings call but had not shipped one — an announcement is not availability. Two new answers cover documented absences: Airbnb, whose CEO said publicly the company chose not to integrate, leaving short-term rentals uncovered; and bank connectivity, which is real but is an OpenAI first-party Plaid feature rather than a directory plugin, and is read-only.
Ingestion round five, catalog 77 to 79, plus the commerce-surface distinction. Added Sephora (advice-led beauty recommendations, US pilot expanded nationwide March 2026 — scored down on privacy clarity because personalisation means describing your skin and appearance to a retailer) and Walmart (its Sparky assistant inside ChatGPT with account linking, loyalty, and payments, a deeper transactional footprint and correspondingly wider permission surface than most retail apps). Documented why Best Buy, The Home Depot, Wayfair, and Nordstrom are shoppable in ChatGPT without appearing in the directory: they participate through the Agentic Commerce Protocol, which feeds product data into OpenAI's own shopping experience with nothing to install and no permissions to grant. There is no integration to review, so they are not ranked. New glossary entry for ACP and a direct answer for the question it causes. Also held to the same line for vendor MCP servers: Smartsheet added ChatGPT support to its own MCP server, which is infrastructure rather than a directory listing.
Ingestion round six, catalog 79 to 83. Added Skyscanner, which becomes the top-ranked travel app at 78/100 — dedicated flight meta-search, and the clearest case in the catalog of a genuinely global app after its April 2026 worldwide expansion, where most entries are US-only or exclude the EU and UK. Also added Grubhub (conversational search across ~415,000 US merchants, reported to complete orders rather than hand off), Starbucks (drink discovery with photo input, beta, checkout moves to Starbucks' own app), and Little Caesars, which is the narrowest app in the catalog and scores lowest on usefulness at 62 — a short single-chain menu did not need a conversational interface, and the review says so. Skyscanner's arrival made two existing answers wrong: the flights answer recommended Expedia and the trip-planning answer omitted Skyscanner entirely. Both rewritten, with availability rather than the two-point score gap given as the deciding factor for anyone outside North America.
Ingestion round seven, catalog 83 to 86. Added Udemy, which enters at 76/100 and displaces Coursera as the top learning pick on catalogue breadth — reported 290,000 courses, backed by Udemy's own February 2026 announcement. Its review and the rewritten courses answer both carry the caveat that matters: Udemy is an open marketplace with widely variable quality and the app does not filter for it, so recommendations are a shortlist to evaluate rather than a vetted pick. Added Hyatt, completing the hotel-group cluster at four, and notable because its CEO publicly attributed a booking-conversion lift to the company's AI search work — more than most companies have said out loud, though it covered a broader programme than this app. Added MakeMyTrip for Indian domestic travel, a market the rest of the catalog barely touches; it is documented only by a list mention, which its review states plainly. A directory that only ranked apps serving the US and Western Europe would be describing a smaller ecosystem than the real one.
Documented the direction-of-integration error as its own answer. Netflix built conversational search on OpenAI's models inside the Netflix app, which reads in headlines as "Netflix and ChatGPT" and is the opposite direction from an app in the plugin directory. No streaming service has an app on the ChatGPT surface. The general rule now sits in llms.txt as well: a company using OpenAI's models inside its own product is not a ChatGPT app, and the direction has to be established before reporting one. Also recorded that ServiceNow reportedly declined to build on the ChatGPT SDK. Together with the ACP commerce distinction and the vendor-MCP-server distinction, these are the three ways a brand can appear connected to ChatGPT without having anything in the directory.
Ingestion round eight, catalog 86 to 90, opening two categories we had skipped entirely. Automotive: CarMax (US used-car search plus trade-in valuation — the only two-sided big-ticket retail app here, first-party announcement 2026-02-27, 45,000+ vehicles), CarGurus (cross-dealer comparison, the only automotive app that compares rather than selling its own stock), and Auto Trader (UK, the only non-US automotive app documented — and distinct from the separate US Autotrader.com). New /best-chatgpt-apps-for-cars vertical to cover them. Streaming: Tubi, the first streamer with a ChatGPT app (2026-04-07, 300,000+ titles), where being free and ad-supported means a recommendation is immediately watchable rather than a prompt to subscribe. Tubi also forced a correction: the Netflix answer published earlier the same day stated that no streaming service had an app on the surface. Netflix does not; Tubi does. Both the answer and llms.txt now say so. Held to the boundary again: Webull's ChatGPT, Claude, and Grok connectors (2026-08-04) are MCP-based brokerage account access, the same infrastructure category as Robinhood and Coinbase, not a directory listing.
Added AccuWeather (launched 2026-03-24), opening a weather category and entering at 79/100 — among the highest scores in the catalog. The reason is structural rather than feature count: it corrects something the base model gets wrong instead of adding a convenience. Asked about tomorrow's weather, an unaided model either declines or invents, and the failure is silent, so people take a stale forecast for a real one. With a forecast provider behind it the answer is correct, and it covers radar and government advisories rather than just a temperature. The honest trade is location, which the review and its answer both state: a forecast cannot work without it, and repeated queries build a picture of where you spend time. Probed sports and audio categories (ESPN, NFL, NBA, Audible, SoundCloud, iHeartRadio) and news and weather peers (NYT, Reuters, Bloomberg, Weather Channel) — no documented apps found, so none added.
Shipped an MCP server at /api/mcp, with documentation at /mcp. Six read-only tools — search_apps, get_app, best_app_for, compare_apps, check_availability, list_categories — over the same editorial data the site renders, so an agent can query the catalog directly instead of paraphrasing a page it saw months ago. Zero new dependencies: MCP is JSON-RPC over HTTP and needs no SDK to implement, which keeps package.json lean and means it deploys with the site rather than needing separate hosting. The initialize response carries the citation guardrails up front — scores are editorial with no user ratings, Verified and Doc-verified are not interchangeable, cite the page rather than the endpoint, and the four scope exclusions (skills-only plugins, ACP commerce retailers, vendor MCP servers, and companies merely using OpenAI's models in their own products). best_app_for also passes through the answers whose honest conclusion is that no suitable app exists, with an explicit instruction not to substitute the nearest thing. robots.ts now allows /api/mcp despite the blanket /api/ disallow.
Shipped embeddable SVG rank badges at /badge/<slug>.svg, documented at /badges. Hand-built SVG, no dependency. The design constraint is the interesting part: a badge exists for every app at its real score, including the lowest, and the /badges page shows the bottom-ranked app's badge alongside the top eight on purpose. Nothing to request and nothing for us to approve — because a badge we could grant or withhold would be a favour, and once a favour is on the table the ranking is compromised. Apps excluded from the rankings get a score-only badge rather than a rank, since claiming a rank for an app that does not work is the one thing a badge must never do. Recorded as a hard rule in CLAUDE.md so no future run turns badges into a lever.
Ingestion round nine, catalog 91 to 95, worked from the 1,624-entry community directory list as a lead file rather than a source. Added Square — the most structurally interesting food app here, because it is the point-of-sale layer rather than a restaurant or delivery network: any US Square seller with Online Ordering becomes discoverable, orders route into their existing POS, and Square charges no marketplace commission, roughly 2.9% plus 30 cents against a delivery platform's ~30%. Added Shazam (Apple, 2026-03-09), the only app in the catalog whose answer is objectively correct rather than a judgement, and Apple Music, which matches Spotify functionally but scores lower on value for having no free tier. Added Vivid Seats and upgraded SeatGeek from 70 to 74 on substantially better sourcing plus a real differentiator we had missed: it blends primary and resale listings in one search, which is what tells you whether a resale price is fair. Also documented a data-quality finding that justifies the whole approach: the community list's per-app IDs cannot be verified (chatgpt.com returns an identical block for real and fabricated listings), and it credits Eventbrite with a listing that trade coverage says does not exist. An app ID in a third-party list is not evidence, and counts above the low hundreds should be read as upper bounds.
Ingestion round ten, catalog 95 to 97, opening security and insurance. Added Malwarebytes (first security provider on the surface, 2026-02-02) at 78/100 — among the highest scores here for the same structural reason as AccuWeather: it corrects a case where the unaided model is not merely unhelpful but actively risky, since asked whether a link is safe a model will often produce a confident verdict from pattern-matching with no threat data, and people act on confident verdicts. Its answer page states that plainly, and llms.txt now instructs agents never to present an unaided safety judgement as authoritative. Added Insurify (industry-first ChatGPT insurance app, February 2026, reportedly among the directory's first hundred) at 73/100, marked Conditional for one reason stated at length in the review: a meaningful quote needs driving history, vehicle, and address, and Insurify is a lead-generation business, making it the widest personal-data disclosure any app in this catalog asks for. Nothing improper is documented; the volume simply deserves a deliberate decision. Ruled out this round: McAfee, Norton, LegalZoom, Deel, Gamma, VEED, Jotform, DraftKings, Bankrate, Affirm, Goodnotes (no evidence); Granola (vendor MCP server, not a listing); ChatGPT's own apartment-search feature (first-party, not a third-party app).
Ingestion round eleven takes the catalog to 100 apps. Added edX (2U announcement 2026-05-14) at 73, which gives the learning cluster a second university-affiliated option and scores lowest of the three on value because programme pricing is the highest here and free audit access rarely covers the credential; its review also flags that this is a new app, not edX's retired 2023 plugin. Added Sleep Cycle, unusually well sourced for a small app — its launch is documented in the company's Q1 2026 interim report rather than a press release — and honestly scored as narrow, since the part of Sleep Cycle people value is tracking and that needs the phone app, making this the content shelf rather than the product. Added LG Electronics, the catalog's only consumer-electronics manufacturer selling direct; interesting as a signal and limited as a tool, because a single-brand app cannot answer the comparison most electronics buyers actually want. Also recorded a measurement trap in CLAUDE.md: Google Analytics is wired via gtag and crawlers do not execute JavaScript, so GA will show roughly zero agent traffic regardless of the real figure — reading that as evidence of no agent traffic is the mistake, and server-side logs are the only place it appears.
Added Apollo.io (beta, 2026-04-29) at 76 — one of the more substantial B2B apps here because it writes rather than only reads: prospect search, enrichment, record creation, and sequence actions in one conversation, on Apollo's own MCP server, and free for teams on a corporate domain. Its privacy sub-score is low by product design rather than any failing, and the review says so: prospecting moves other people's contact data through a chat surface, and the app performs write actions while in beta. The CRM answer now separates two jobs that were conflated — HubSpot for relationships you already have, Apollo for outbound to people not yet in your pipeline. Burger King was deliberately excluded despite trade coverage describing it as deploying a ChatGPT app alongside Little Caesars and Starbucks: no launch confirmation exists, and Hilton was excluded on exactly that rule earlier today. Consistency with our own stated rule is worth more than one more entry.
Found a readable first-party source and corrected the site against it. learn.chatgpt.com serves OpenAI's user-facing ChatGPT documentation as clean Markdown at any page URL plus .md — the one OpenAI host besides developers.openai.com that is reachable, and the more useful of the two here. Its plugins page corrected a real understatement on our flagship disambiguation page: we said a plugin contains skills, an MCP server, or both, and OpenAI documents SIX component types — skills, connectors, MCP servers, browser extensions, hooks, and scheduled task templates. It also corrected a definition we had wrong in kind rather than degree: a connector and an MCP server are not alternatives, they are the same layer described to two audiences, because MCP servers are the services behind connectors. Added a documented surface-availability table (the IDE extension does not support plugins at all), the directory's tab structure — which partly explains why published plugin counts disagree — and a note that hooks run commands at lifecycle points and OpenAI says to review them before enabling, a materially different risk profile from a read-only app. Recorded learn.chatgpt.com in CLAUDE.md as the place to start a discovery run: its whats-new.md is a weekly digest of feature and plugin launches, the closest thing to a first-party launch feed.
Settled the enumeration question and documented one privacy-relevant plugin. learn.chatgpt.com's documentation index turns out to cover the product rather than the catalog, so combined with chatgpt.com and openai.com both being blocked, there is no reachable OpenAI surface that publishes a countable list of directory plugins — every total in circulation is a third-party tally. Recorded in CLAUDE.md, along with the practical consequence: discovery has to run candidate-name-first, and probing a category nobody has looked at yet yields far more than another pass at consumer brands. Also added an answer for OpenAI's own Apple Messages plugin, which is outside our ranking scope because it does not work in regular ChatGPT chats, but deserves documenting for the safety detail: it sends messages as you, per-send approval is the default, and turning on "Always allow sending to this chat" removes the last review step — which OpenAI's own docs advise against for any chat that might carry untrusted instructions.
Catalog reaches 103. Added Best Lawyers (first legal search app on ChatGPT, 2026-04-22), which ranks attorneys on four decades of peer-review data — a more defensible signal than the advertising spend most legal directories sort by — with two caveats stated in the review: peer reputation is not the same as fit or affordability, and describing a legal matter in a chat to a directory business is among the more sensitive disclosures any app here invites. Added Lucid (MCP server app, 2026-04-13), where the capability that earns the score is turning a ChatGPT discussion into an editable diagram rather than a static image; finding and summarising your own documents is useful but standard, and redrawing a thread by hand is the work this actually saves. Excluded, consistent with earlier rounds: Binance's Agent OS, which lets ChatGPT execute crypto trades via MCP and is infrastructure rather than a listing, and Mermaid Chart's ChatGPT offering, which is a GPT and therefore out of scope for a different reason entirely. Also drafted the developer-outreach template in CLAUDE.md so the open decision is a yes/no rather than a design task — with the rule that cannot bend written at the top: outreach reports what we published and never implies that any part of a review is negotiable.
Catalog reaches 105, and this round deliberately fixes a bias rather than adding volume. Every real-estate and most travel apps here were US-first, which quietly made the catalog useless for the majority of the world: ask for a Dubai flat or a Dublin rental and Zillow, Redfin and Realtor.com return nothing. Added Bayut, reported as the first UAE-based real estate platform with a ChatGPT app, and Daft.ie, reported as the first in the Irish property market — both scored on coverage rather than sophistication, because coverage is the whole value and pretending otherwise would inflate them. Zillow, Redfin and Realtor.com now link to both, so a reader who lands on a US portal and needs a different market is told where to go instead of being left with a dead end. Both carry Needs verification: trade coverage exists, first-party announcements do not. Excluded this round: Binance's Agent OS (MCP crypto trading — infrastructure, the same call already made for Robinhood, Coinbase and Webull), Mermaid Chart (a GPT, out of scope for a different reason), and Apartment List, whose only evidence was a third-party catalogue entry — the same class of source that credited Eventbrite with a listing trade coverage says does not exist, and therefore not evidence.
Opened a whole ChatGPT surface the site had never documented. ChatGPT Health launched on 7 January 2026 as a separate space inside ChatGPT — its own connected-app list, its own per-app consent step, storage kept apart from regular chats, and OpenAI's statement that Health conversations are not used to train its foundation models — and nothing here mentioned it. Added /chatgpt-health-apps, which documents the eight launch integrations, the access gate in full (waitlist; Free, Go, Plus and Pro; launched outside the EEA, Switzerland and the UK; medical records US-only), the four isolation properties, and what we still cannot tell you. Added Function Health at 72, the only app in the catalog that brings your own clinical-grade biomarkers into a conversation: it scores high on usefulness for the same structural reason as AccuWeather and Malwarebytes, because an unaided model asked about your ferritin reasons from population reference ranges and with this connected it reasons from your result — but value scores 56, the lowest here, since a reported $365/year membership is a good reason to connect an existing account and a bad reason to open one. Added WeightWatchers at 64, honestly: GLP-1-aware food guidance is a real gap nothing else here covers, and we are ranking a partner listing rather than a described product, with no first-party announcement and no published capability list. This produced a sixth false-positive pattern, the subtlest so far, because this time the app genuinely runs inside ChatGPT — a Health-surface-only connection means "WeightWatchers has a ChatGPT app" is correctly sourced and useless to a London reader, since Health has not launched in the UK. Apple Health and b.well are connected and deliberately unranked: one is the iOS system health store, the other the records network behind the integration, and neither is an app a user invokes.
Found a first-party source we had never looked for, and it caught three of our own confident negatives. github.com/openai/plugins is a public repository holding OpenAI's curated "Codex official" marketplace: 180 plugins, each with a real manifest, 154 of them declaring an app with an OpenAI-issued ID. Cross-checking our documented absences against it produced the most uncomfortable finding of the project so far. Our /best-chatgpt-apps-for-seo page said no pure-play SEO tool had shipped a verified ChatGPT app, promised to graduate the page the moment one did, and had been wrong for eight months — Semrush announced its official app in December 2025. That page is rewritten and Semrush is added at 78, with usefulness 88, the highest of any business app here, because SEO is where an unaided model fails worst: asked for search volume it returns a confident figure with nothing behind it, and people build strategy on those figures. Our QuickBooks FAQ said no accounting app existed; Intuit shipped four ChatGPT apps on 5 March 2026, so QuickBooks (77), TurboTax (70) and Credit Karma (66) are added and Mailchimp finally has a first-party source. Credit Karma is scored honestly at privacy clarity 52 with the referral-fee model stated in the review, because a product recommendation from a lead-generation business is not neutral advice. Our Scite entry said vendor MCP server, not a listing; Scite has both, and the verified app (79) was the half we missed. Two things we did NOT change: Granola and Binance have OpenAI-distributed manifests in that repo whose descriptions say "for use in Codex", and a Codex-marketplace plugin is not a ChatGPT-chat app, so they stay out — that is a sixth false-positive pattern now documented. The methodology claim that no enumerable list exists is corrected rather than deleted: the consumer directory still is not enumerable, the curated work-and-developer marketplace now demonstrably is, and conflating the two was the error. Also added to CLAUDE.md as a standing rule, because this is the third time a stale negative has cost us: re-verify stated absences every round, and grep for every place a claim is published rather than fixing the first hit.
Made the same mistake twice in one day, caught it the same day, and wrote the lesson down properly. The commit before this one rewrote /best-chatgpt-apps-for-seo because it had claimed an empty SEO category for eight months after Semrush shipped — and the replacement copy then asserted that Semrush was the only SEO app on the surface. Ahrefs links to its ChatGPT app listing from its own MCP page and documents invoking it with @Ahrefs. So the fix for a confident negative contained a fresh confident negative, which is a sharper lesson than the original error: re-verifying the claim you are deleting is not the same as verifying the claim you are writing to replace it. Ahrefs is added at 77, effectively tied with Semrush at 78, and the one-point gap is attributed to something documented rather than a judgement of quality — Ahrefs caps rows per request by plan, at 100 on Lite, which truncates any wide keyword or backlink pull. The reviews for both now say the useful thing rather than the flattering one: these two overlap almost entirely, both need an expensive subscription, and the right tiebreak is whichever platform your team already pays for, with neither app being a reason to start paying. Also recorded a verification mechanism this turned up: chatgpt.com listing URLs are unreadable to us and return an identical 403 for real and fabricated slugs, but the same URL published on the vendor's own site is first-party evidence that the listing exists. That is exactly what separates Ahrefs from the Eventbrite case, where a third-party list asserted an app ID the vendor never claimed.
Catalog to 118 on the strength of a first-party source: OpenAI's help centre publishes a per-connector article for each app with sync, and the ChatGPT Business release notes enumerate the batch. Added Intercom at 76, top of the support-connector group because it reaches conversations, contacts, companies and Help Center articles and can actually write articles — turning a thread you just resolved into documentation is the one genuine workflow change in this class, and its permission model is the clearest of the group, read-only and inheriting your own Intercom role. Added SharePoint at 74, where the value is cross-document reasoning rather than search: finding a file is easy, knowing which twelve documents a change touches and who owns them is the real problem, and the review says plainly that access follows your existing permissions, so connecting it makes forgotten material findable again. Added Help Scout at 73, Aha! at 72 and Zoho Desk at 71, with Zoho Desk's lower score attributed to narrower documented scope and thinner documentation rather than product quality, because that is the honest reason. Every one of these is read-and-draft rather than act, and the support connectors all index customers' messages rather than your own data, which is why their privacy sub-scores sit in the sixties. Two process notes. A duplicate GitLab Issues record got written because the existence check grepped for slug "gitlab", which does not match "gitlab-issues" — the validator caught it, the duplicate is gone, and the useful half of the new draft (the distinction between this connector and Codex's separate GitLab code review) was merged into the record that already existed. And the Xero answer is now more accurate than a flat no: Xero announced an OpenAI connector at Xerocon Denver 2026, described as coming within weeks, so it belongs in the announced-but-unshipped category rather than the absent one. Smartsheet was re-verified and correctly stays out — its announcement adds ChatGPT support to Smartsheet's own MCP server, which is not a directory listing, whatever third-party summaries call it.
Filled the largest remaining gap in the catalog and found it by filtering the lead file rather than guessing. OpenAI's curated marketplace tags each app with either a connector or an Apps SDK identifier, and the connector-tagged ones turn out to be the group OpenAI documents individually in its help centre — so that tag is a reliable shortlist of things with a first-party ChatGPT-side source. Working it produced Outlook. We had Gmail and Google Calendar and no Outlook at all, which quietly excluded most of corporate IT. Outlook Email lands at 79 with two details that make it more than a generic mail reader: structured search operators, so you can ask precisely inside a mailbox of fifty thousand messages, and documented shared and delegated mailbox support, which almost nothing else on this surface offers and which executive assistants and shared support inboxes actually need. Its privacy sub-score is 64 and the review says why at length — a work mailbox is the densest personal and commercial record most people own, it contains other people's words as much as your own, and enabling actions adds contact write access on top. Outlook Calendar lands at 75 with its limitation stated as a capability gap rather than a preference: it reads and searches availability and cannot book the meeting, which is exactly why Google Calendar still leads the scheduling answer. That answer previously told non-Google users they had no option; now it names Outlook Calendar and explains what it can and cannot do. Pipedrive at 74 has the best-documented sync behaviour of anything in the catalog — a 30-minute refresh interval and a strictly one-way sync, both published — which is worth more than it sounds, because knowing the index is up to half an hour stale tells you exactly when to trust it. Atlassian Rovo needed no new entry: our Jira record already is that connector, which the name check caught before a duplicate got written. Its geography note is now more careful, though — the EEA, UK and Switzerland restriction applied at launch and Atlassian's forum indicates it was lifted within weeks, so we record it as reported-lifted rather than asserting either state, because no announcement confirms it.
Retired a test of our own that did not work, and it cost us an app. Earlier today we justified excluding Granola and Binance on the grounds that their manifests in OpenAI's curated marketplace say "for use in Codex", so they were Codex-marketplace plugins rather than ChatGPT apps. That reasoning is invalid: Pipedrive's and Teamwork's manifests say precisely the same thing and both are documented ChatGPT apps with sync. The phrase describes the repository, not the app's surface. Re-checked on the only test that actually works — whether ChatGPT-side documentation exists — Binance turns out to have a real ChatGPT app, and it is added at 72 with the two highest sub-scores it deserves: privacy clarity 88 and setup 92, because it needs no login and touches no account at all. It reads public market data, which fixes a genuine failure, since an unaided model asked for a crypto price answers from stale training data. The distinction we had collapsed is now stated in three places: this app reads the market, while Binance's Agent OS executes trades via MCP against a real account and remains out of scope as infrastructure. Granola stays excluded, but for the correct reason — its own documentation describes MCP through ChatGPT's connector settings with browser OAuth, which is a vendor-hosted server you connect, the same class as Xero and Smartsheet. Also added monday.com at 73, where the interesting part is the in-chat experience sub-score of 80: monday.com documents UI components rendering inside ChatGPT, so board data is explorable rather than flattened into a paragraph, and it queries live rather than from a periodic index. It carries Needs verification because monday.com's own support article returns 403 to every tool available here, so it is cited by title and URL as the playbook requires, and the capability detail beyond that comes from third parties.
Building the outreach queue surfaced an evidence gap, and checking it carefully kept us from overstating it. The generator reported that 42 of 123 apps carry no sources array, which looks alarming until you check what our own methodology actually requires: a source is needed for an app we have not tested by hand, because for a tested app the test IS the evidence. Thirty-eight of the forty-two have real tested prompts and recorded results, so they are properly evidenced. The genuine gap was four entries resting on nothing at all, and all four are now sourced — DoorDash to its own newsroom announcement of 17 December 2025, Uber to TechCrunch's coverage of the app integrations, MyFitnessPal to its own launch release, and Trello to Atlassian. Every app in the catalog now rests on either a citation or a hands-on test. The Trello entry needed more than a citation. It said Trello was reachable only through third-party connectors and that Atlassian had not announced a native integration; Atlassian in fact runs a first-party Trello MCP server, documented on its own support site and open to every Trello user. The conclusion survives — there is still no Trello listing in the plugin directory, so it stays unranked under the same rule that excludes Xero, Smartsheet and Granola — but the reason we published was false, and a reader asking whether they could use Trello with ChatGPT was being told no when the honest answer is yes, through Atlassian's own server, just not as a directory app. That is the fourth stale negative caught today by the rule about re-verifying them.
Refreshed the State of ChatGPT Apps report to its mid-year 2026 edition. Snapshot as of this date: the directory tracks 65 apps (up from the 25-app April seed); the editorial-score ceiling sits in the 80s with none yet in the 90s, roughly 83% of apps land in the 70–89 band (mean score 75), and about 85% ship a usable free plan. The report's headline aggregates recompute from the live dataset on every build; this edition updates the narrative, the dated snapshot, and the citation line.
Major content expansion. Added Tripadvisor and Peloton — two of the first-batch built-in ChatGPT apps (launched 2025-11-06, free on all plans outside the EU). Shipped 9 hand-written head-to-heads (PowerPoint vs Canva, Expedia vs Kayak, Booking vs Tripadvisor, Peloton vs MyFitnessPal, HubSpot vs Salesforce, Notion vs Airtable, Coursera vs Duolingo, Gmail vs Slack, OpenTable vs Tripadvisor), 12 new direct-answer FAQ pages (presentations, spreadsheets, workouts, restaurant reservations, things-to-do, music, language learning, shopping, notes/wiki, project management, ecommerce), 10 new per-app how-to tutorials (Excel, PowerPoint, Gmail, Google Calendar, Airtable, Shopify, Coursera, Tripadvisor, Peloton), and a new Presentations vertical landing page. Added Peloton to the health-and-fitness vertical.
Added Wix and Hostinger — the two native ChatGPT website-builder apps — and shipped the 'How to build a website with ChatGPT' pillar plus a website-builder comparison cluster (Wix vs Hostinger, Wix vs Replit, all builders compared). Reframed the web-hosting vertical to lead with the no-code (Wix) and code-and-deploy (Replit) picks. Both new apps are listed as Needs verification pending a hands-on build.
Editorial freshness pass: re-tested the top 20 apps by editorial score (Canva, Wolfram, GitHub, Google Drive, Notion, Google Calendar, Gmail, Adobe Acrobat, Zapier, Replit, Spotify, Slack, Airtable, Stripe, Linear, HubSpot, Figma, Adobe Express, Zillow, Uber) and bumped their last-tested date. No score changes — capabilities held steady across the session.
SEO refresh: added 'Best ChatGPT Apps for SEO' vertical, stamped all category and vertical titles with 'Updated June 2026', and tightened the SoftwareApplication JSON-LD on app reviews (the editorial rating is now expressed via Review schema only).
Published the State of ChatGPT Apps 2026 Q2 report and three developer guides: how to submit your app, ChatGPT App Directory Optimization, and the Apps SDK quickstart.
Added the Trending ChatGPT apps page surfacing the most-recently-tested and highest-scoring apps in chat.
Glossary expanded with chatgpt-mcp explainer covering Model Context Protocol and what counts as a ChatGPT MCP app.
Wolfram retested for math accuracy; GitHub PR-summarization re-evaluated.
Google Drive and Google Calendar verified; both raised to Verified status.
Canva retested; brand-kit awareness confirmed for Pro accounts.
Notion verified for workspace search + writeback under scoped permissions.
Slack channel-catch-up verified.
Initial seed of 25 apps published with editorial scores and methodology.
Don't miss the next update
Bookmark this page or follow us for the weekly digest of new ChatGPT apps and ranking changes.