Screen shares carry their sound, calls are loud enough to hear, and a stopped share no longer leaves everyone staring at a black screen. Attach a Codex page to the AI the same way you attach a file, and the document formats that quietly failed on the server now read everywhere.
New5
Share your screen with its soundA shared tab or window now carries its audio, so a video, a demo or a recording is heard by everyone instead of played in mime. Music and soundtracks are sent at full stereo quality rather than as speech, and your local screen recording captures the sound too.
Change your microphone, speaker or camera mid-callA small arrow on the Mute and Camera buttons opens the device list, so you can pick a headset you plugged in after joining, switch to the other camera, or send the sound to a different speaker without leaving and rejoining. Devices that appear or disappear during the call update the list on their own.
Calendar privacy is a setting, not a ruleA switch under Settings, Members and permissions controls whether admins can see more than Busy on events they are not part of. It is off by default, so an admin sees exactly what everyone else sees, and only someone who can manage the workspace sees the switch or is affected by it.
Attach a Codex pagePick pages from your workspace and the AI reads them in full, keeping headings, checklists, tables and spreadsheet cells rather than flattening everything into plain text. Search across your pages, tick as many as you need, and each one attaches as a snapshot of the version you chose.
One menu for everything you attachThe plus button now offers files and images, or pages from your workspace. Attached pages open in the side panel exactly like a file, so you can see what the AI was given.
Improved2
Meetings are loud enough to hearCalls came through too quietly, and turning your volume up did not help because the meeting was already playing at full. Everyone now comes through louder and the quietest people gain the most, so a colleague on a laptop microphone across the room rises to meet someone on a headset.
More formats, and fasterDatabases, GIF screenshots and archives read for the first time, and Word, Excel and PowerPoint keep working even if a reader is unavailable. A short PDF that used to spend 31 seconds in image recognition now opens in under a second.
Fixed7
A stopped screen share no longer leaves a black screenWhen you stopped presenting, your own window cleared but everybody else kept a full size black rectangle labelled Presenting for the rest of the call, with nothing to click to remove it. The share now disappears for everyone the moment it ends, and the call re-checks itself so it can never stay stuck.
Everyone sees the same meeting layoutWith two people in a call the tiles could be laid out as two full width bands that cut off the top and bottom of the picture, and which one you got depended on the exact size of your window, so the same call looked right for one person and broken for the other. Two people now sit side by side, and no tile can be stretched wider than the picture it holds.
Blocked meeting audio can be switched back onBrowsers refuse to play sound that nobody has clicked to allow, and when that happened the call was simply silent with nothing to say why. There is now a button that turns it on.
Vernais looks like itself everywhereOutside Wire the assistant was drawn with a generated human face, so the Team list, the activity feed and its own profile card showed a stranger. It now carries its own mark on every surface, sits in line with everyone else in a conversation, and its profile no longer reports it as offline.
PDFs and spreadsheets open againThe document readers were missing from the server, so Word, Excel, PowerPoint and PDF attachments came back empty and the message blamed your file. Every document format now reads on a clean install, and coverage went from 26 of 35 test files to 34.
No more guessing at a file it could not readWhen an attachment failed, the AI was being asked to describe the file from its name, so a CV it never opened came back summarised. It now says plainly that it could not read the file, and gives the real reason.
A missing reader is named as a server problemA file the server cannot open is reported as a server problem with the fix, not as a broken document. Older Word and PowerPoint files now say which format to save as instead of returning a parser error.
Vernais now works wherever you do: tag it in team chat, hand it any file, ask it to write a page. Initiatives finally carry the whole investigation instead of losing it at creation.
New7
Vernais in WireTag the AI in any channel or direct message and it answers inline, reading the thread and your connected data. Read only, under the permissions of whoever asked.
Conversations into initiativesAsk Vernais to turn a Wire discussion into initiatives. It splits the conversation into distinct topics and shows a card with checkboxes. Nothing is written until you tick and click.
Build your own chartsA chart explorer over your knowledge graph: 329 fields across 24 tables, drag onto axes and filters, 21 chart types plus KPI tiles, tables and pivots. Save into cross-filtering dashboards, export to CSV.
Attach almost any file21 formats verified readable, up from images and plain text. Photos and charts are described, screenshots are read at 99.1% character accuracy, and every file opens in a side panel showing the exact text the AI was given.
Ask for a document or a web pageVernais drafts a full Codex page, or a self-contained web page with its own styling. Both render live, you revise by asking, one click publishes.
Spreadsheets and real block editing in CodexA spreadsheet block with 500+ formulas, sort, filter and tabs. Select all escalates from text to block to page, and copy, cut and paste keep a block type instead of flattening it.
A sample dataset in one clickImport a demo company across every connector with the graph already built, so Data and investigations have something to show immediately. It never touches your own chat, pages, calendar or connections.
Improved5
Far more of the right evidenceRetrieval picks a strategy per question instead of always taking the top matches. On an independent benchmark, evidence recall went from 42.5% to 86.2%.
Web search leads with authorityFirst-party sources rank separately from topical relevance, and disagreement between sources is surfaced. Authoritative results in the top five went from 44% to 63%, and content-farm results went from 4 to 0.
Measure on your cadenceChoose how often a locked metric is read. Each rate carries a confidence interval, and a move must clear both a statistical and a practical bar before it counts as improvement.
Faster to openA hard-coded five second loading wait is gone, first load downloads 33% less code, and creating a workspace takes about 2 seconds instead of 40.
Renamed sectionsLive chat is now Wire, Pages is Codex, and Calendar is Beam. Your conversations, pages and events are untouched.
Fixed7
Initiatives keep the investigationEvidence rows read "Grounded fact cited in the investigation" with nothing behind them, and event counts read 0. The investigation is now captured whole at Generate.
Evidence tagged later actually countsA proposal tagged from chat, Wire, a page or a transcript now lands in Validate or Decide and moves the confidence, instead of being discarded.
Confidence stopped reading 76%Every initiative showed exactly 76%. The score tracks the verified evidence pass rate again, so the number carries information.
Screen share keeps your cameraSharing replaced the presenter tile, so your face vanished and did not return. Each camera and screen is now its own tile, with Auto, Grid and Speaker layouts.
Opening an initiative is instantDecide made a live AI call on every page open because a cached verdict read as a miss. That went from 6.95s to 0.05s.
Renaming yourself applies everywhereOld messages, comments and invites kept your previous name and avatar. Both now resolve fresh on every read.
Stopping a scrape stops itStop left the crawl and its browser running, and one conversation cleanup could kill another conversation scrape.
The AI chat was rebuilt from scratch, and the investigator behind it became a tool-using agent you can watch work.
New6
A completely rebuilt AI chatPersistent threads, expandable live reasoning, file attachments, and far better markdown, table and code rendering.
Watch the AI investigateThe investigator decides what to look at next instead of following a fixed script, streaming a live timeline into chat. Its toolkit grew from 11 to 15 tools, including tracing the real connection between two records.
Investigation to initiative in one clickGenerate a full initiative, with problem statement, signal clusters, hypotheses and proposed solutions, without re-running anything. Over 10 minutes down to 15 to 40 seconds.
Split one investigation into severalWhen an investigation surfaces more than one distinct problem, you can create them separately, merge them, or group them your way.
Coverage checkEvery investigation estimates how much of the relevant evidence it actually found, and flags the answer as thin when it did not find enough.
AI inside a pageRewrite, continue or generate a block from the block menu, generate rows for an empty table, and diff two versions before restoring.
Improved5
Graph rebuild 19x faster and repeatableA full rebuild of the same data dropped from 417s to 21.5s, and two runs on identical data now produce an identical graph, which was never true before. A silent 20,000 row limit that left large workspaces quietly building an incomplete graph is gone.
One Filter control everywhereSeven sections lost their crowded left rails for a single Filter dropdown, and Settings became its own full-screen destination.
Data shows the whole graphA Max nodes control with presets or All, replacing a hard 5,000 node cap that hid the rest of a 20,000 node workspace.
Storage works like a real driveWord, PowerPoint and Excel preview in the browser, search reads inside documents, and you can multi-select to copy, move or delete.
Richer executionFormatted task descriptions, each selected solution getting its own task group, and a board that loads in 15ms instead of 5,027ms.
Fixed7
The answer is no longer thrown awayLong investigations showed "Investigation complete" and then "synthesis unavailable". The real answer had been written and discarded because an internal timer fired mid-composition.
Take a reading returns real numbersReadings reported 0 for three different reasons that all looked identical on screen. It now reads real telemetry, and when it genuinely cannot, it says why.
Data filters search everythingFilters applied to only the first few thousand rows while the badges counted the whole graph, so a filter could return an empty table.
Drag and drop sticksMoving a card or changing a stage often appeared to do nothing. A request already in flight answered with the old state about 62% of the time.
Confidence agrees with itselfA high verdict could render as 65%, because the list row and the detail page computed confidence two different ways.
Records read like recordsOpening a record showed a REST URL and a blob of JSON. It now shows the ticket reference people quote, plus status, priority, assignee and type.
The grey application error screen is goneAny single failure blanked the whole app. There is now a real error screen with retry, and a stale tab after a deploy heals itself.
Collaborative pages, real video meetings that transcribe themselves, and bookmarkable URLs everywhere, with a nine round security audit underneath.
New7
Live collaborative editingMultiple people write in the same Codex page at once, edits merging character by character, each teammate caret shown as a named coloured line. One Share control covers Private, Workspace can view, and Workspace can edit.
Page version historyAutomatic checkpoints roughly every 45 seconds of real editing, with a read-only preview, a block-level diff and undoable restore.
Meetings that transcribe themselvesCalls run on our own service, and a meeting can be transcribed for the whole room rather than just whoever has the tab open, saved per speaker into the meeting transcript page.
Invite codes and company accountsAccess is invite-gated with single-use codes, and colleagues on the same verified company domain join the shared plan instead of each creating a parallel subscription.
Bookmarkable URLsEvery section and the things inside it have a real URL. Back and forward work, and a refresh no longer drops you on the dashboard.
Export your workData to CSV, Excel or JSON. A Codex page to Markdown, HTML or PDF including tables. The activity log honouring the active filter.
Automatic per-connection syncSet a connection to sync itself every 15 minutes, hour, 6 hours or day. An unchanged page costs nothing and creates no duplicates.
Improved5
Chat answers in about a secondThe extra model call before every message is gone and answers stream as they are written. A reply starts in about 0.8s, instead of 1 to 3s with occasional 20s stalls.
Your data is locked downA nine round security audit closed over 145 findings: cross-workspace leaks, ID guessing on files, unsafe outbound requests, session and cookie hardening, and rate limiting.
The call is a real windowDrag, resize, minimise to a picture in picture pill or go fullscreen, and keep using the app mid-call. A green room lets you check camera and mic before joining.
Readable knowledge graphThe graph laid thousands of nodes out with spacing tuned for dozens. On a real 4,851 node graph, overlapping nodes went from 205 pairs to zero.
Pages load faster under loadNine parts of the app opened a new database connection per query. Pooling took a representative call from 2,009ms to 13ms under 32 concurrent requests.
Fixed5
The AI says when it cannot prove somethingAn investigation only states a cause when the evidence chain holds up. Otherwise it returns a clearly labelled hypothesis and names the measurement it is missing.
Wire messages arrive instantlyIn production the live connection was rejected before it opened, so new messages only appeared when you switched sections.
Signing in lands you in the right accountThree compounding problems meant you could sign in as yourself and open somebody else workspace.
Verification and reset emails arriveBranded emails were accepted by the relay and then dropped before the inbox, because two required headers were missing.
Storage downloads work in productionDownloads and inline previews that failed with "site can not be reached" are fixed, alongside single click folders and a right click menu.
Every section re-authored on one design system, landing alongside downloadable reports, per-person permissions, and an execution phase that reaches into real code and then watches production for the regression it caused.
New7
Downloadable initiative reportsEvery initiative auto-generates a report covering the finding, evidence, hypotheses, decision, metrics and sources, downloadable as a designed PDF with a cover, contents and charts.
Per-person permissionsThe owner ticks each capability for each teammate and their navigation reacts immediately. A section they lose access to disappears without a reload.
Branches, pull requests and merges from a taskCreate a branch from a task, watch its commits, pull requests and build checks appear automatically, then merge from inside the task.
Regressions file themselvesAfter a merge the task enters a soak window where your error data is scanned and scored against that branch, commits and release. A confident match is auto-filed as a regression bug.
Design your own workflowsDrag states around a canvas, draw arrows between them, rename stages and set a workflow live. Board columns and allowed moves follow it.
Command paletteOne keystroke searches initiatives, tasks, products, pages, meetings, team members and knowledge graph records in one place.
Five new tabs per connectionEvents, Logs, Webhooks with a unique inbound URL and signing secret, plus Permissions and Settings.
Improved6
One design system across 19 sectionsConsistent cards, hairline tables, per-section sub-navigation and tuned dark and light themes. Behaviour and data are unchanged.
The graph updates itselfRecords added by a sync, and pages from a web scrape, join the graph without pressing Run Corvex, matched against everything already there.
One assistant, no routing delayChat no longer runs a separate classifier before answering. Every message goes into a single loop that decides for itself which tool to use.
Measure tracks a set of metricsA north star metric plus supporting metrics, each with its own formula, receipt, baseline and change.
Permissions enforced, not just hiddenEvery mutating action checks the caller permissions on the server too: integrations, graph rebuild, investigations, billing, team and reports.
New Vernais brandingA new wordmark and mark throughout, with real photographic artwork on sign-in and onboarding.
Fixed3
Account takeover and privilege escalation closedPassword reset and invite tokens could be returned to unauthenticated callers on a particular deploy configuration, and a user could escalate themselves to platform administrator through the email change flow.
Chat and web search after a provider outageOne model provider started timing out on every call, silently breaking answers and search. Requests now fall back transparently.
Investigations no longer look stuckA running investigation collapsed all its stages into a single row, so it appeared frozen for minutes.
Execution became AI-planned, and a new Measure phase plus a Launches section answer the question the workflow never could: did the thing you shipped actually work?
New8
Measure, the fifth phaseEvery investigation derives one tracking metric with the full formula and the counts behind it. Lock that value as a baseline, take readings, and mark the work shipped, paused or rolled back.
LaunchesA post-ship outcomes view listing every initiative metric as an improving or regressing trend card, with breakdowns by owner, impact and product.
The AI plans the workMoving to execution hands the chosen solution to the AI, which reads the investigation evidence and breaks it into the minimum complete set of tasks. You review and approve before anything is created.
Engineering tasks arrive with the code locatedA technical task attaches the concrete issues to address plus the specific files and line numbers where each one lives.
StorageA file browser over everything the product produces or you upload: recordings, initiative files, reports, chat attachments and scrapes, with auto-organised folders, search, preview and a live quota meter.
DashboardA live overview built from your own data: signal volume, source health, average confidence, decision-ready initiatives and where your team time is going.
My tasksA personal queue gathering everything assigned to you across every initiative, with board, list and table views.
Readable IDsIN-12 for an initiative, IN-12-H1 for a hypothesis, IN-12-S1 for a solution and IN-12-T1 for a task, shown everywhere.
Improved5
Data is the single knowledge hubThe separate Graph and Sources sections folded into Data, with one filter rail driving a Table and Graph toggle plus a Scraped tab.
Hypotheses show their evidenceEach opens onto the cited records with source and timestamp, a step by step causal chain, and a confidence broken into the six factors that produced it.
Faster web answersSearches for one question run in parallel rather than sequentially. Six searches went from 23.1s to 4.8s.
Click any numberSignal counts, evidence factors and cluster figures across Identify and Validate open the Data browser filtered to exactly those records.
Integration credentials encrypted at restAPI keys and tokens are encrypted before storage, and sensitive endpoints are rate limited.
Fixed3
Clusters stopped filling with noiseIdentify clustered the investigation broad measurement scope, so a checkout investigation showed clusters of unrelated repository activity. It now clusters only the grounded evidence.
Chat answers general questions againAsking about a definition, a concept or a public product specs could return a refusal, sometimes after it had already searched the web.
Leaving chat mid-answerSwitching sections while an answer was streaming destroyed the whole chat and killed the live progress. It now keeps running in the background.
Vernais stopped being a single-user tool. Real accounts, isolated workspaces and billing landed first, then the sections that turn an AI root-cause answer into scheduled work.
New12
Accounts and isolated workspacesEmail and password sign-in with secure sessions, plus workspaces you create or join by invite code. Each workspace gets its own isolated database, so chat, knowledge graph and AI memory never leak between them.
Sign in with GoogleA Continue with Google option next to email sign-in, with existing accounts linkable.
InitiativesTurn an investigation into a tracked initiative in one click, carrying the problem statement, root cause, confidence and evidence. It moves through Identify, Validate, Decide and Act with competing hypotheses, ranked solutions and a scheduling board.
Evidence from everywhereTag an initiative from a Wire message, a meeting transcript or a Codex page and it becomes a tracked signal, cleaned up, judged for relevance and sorted, with repeats merged rather than duplicated.
WireReal-time team chat: channels in categories, private channels, direct and group messages, forwarding, replies, edits, pins, reactions, mentions, typing indicators, file uploads and search. Encrypted at rest, with optional end-to-end encryption for one to one messages.
CodexA block editor with a slash menu, headings, to-dos, tables, code, callouts, toggles, mentions, comments, templates and databases with formulas. Import a PDF, Word, Excel or PowerPoint file straight into a page as editable content.
BeamMonth, week and day calendar views with guests, reminders, repeat rules and overlapping events laid out side by side, plus availability and inline scheduling. Guests can propose an alternative time.
Video meetings that transcribeCalls run in the app with a participant grid, screen share, chat and live captions. With recording on, the transcript is filed as a dated page and as evidence on the linked initiative.
Data, Inbox and Activity logA full-width browser over every node in your knowledge graph with search, filters and a detail drawer. Meeting and invite alerts that refresh on their own. A time-stamped feed of every meaningful action.
Settings, plans and usageNine panes covering profile, security and sessions, notifications, workspace, members and permissions, invitations, plan and appearance, with live usage meters and enforced workspace and seat caps.
Web research with confirmationSay what you want researched and the AI walks you through page count and depth, finds the sources, flags which are paywalled and asks before including them, then scrapes with a live progress panel you can stop.
Admin portalA separate back office to search and inspect every account, manage plans, discounts, credits and limits, and read an append-only audit trail with before and after diffs.
Improved3
Attached documents read properlyScanned PDFs, images, Word, Excel, PowerPoint, HTML and CSV extract cleanly, keeping tables and headings, and stay in context for the whole conversation.
A Sources panel for web answersClicking the sources pill opens a side panel of source cards with favicon, publisher, title and snippet.
Real syncing across every connectorLive endpoints, pagination styles, primary keys and cursors now drive syncing for every supported integration.
Chat grew a real conversation workspace and learned to answer from your connected data, from a live web search, or from a full investigation.
New6
Saved conversation threadsStart a new chat, search past conversations, rename or delete them, and jump back into any thread with its full history.
Live web search with citationsAsk about something outside your workspace and chat runs real searches, opens the best pages, and answers with inline citations plus a live timeline of every query and page read.
Grounded data lookupsAsk about your connected data and get a count of matching records with the rows behind it, computed from the records rather than generated by the model.
File attachmentsAttach PDFs, Word, Excel, PowerPoint, CSVs, code, logs and images up to 25 MB and ask about them.
Connect, test and sync your toolsEnter credentials, run a one click connection test, pick which streams to sync, and drill into a searchable field tree with fill rate bars to select only the fields you want.
Step by step setup guidesEach integration has an in-product guide: a direct link to the right settings page, the exact steps to create the credential, and the scopes to grant.
Improved3
Every investigation number carries its filterCounts, breakdowns and samples come from tools that return the exact filter behind each figure, so a number can be reproduced. Counts were verified exact against a 917,623 row workspace.
Success and failure mean the right thing per toolOutcome words are almost never literal field values, so the AI discovers what success and failure mean for each tool before counting.
Measured impact counts both sidesImpact numbers were failure-biased, reporting 7 successes when checkout had 138. A checkout investigation now reports 427 events, 224 failed and 138 succeeded, matching a direct data lookup.
Fixed3
Investigation verdicts stopped contradicting themselvesThe headline verdict came from an earlier draft the report sometimes went on to reject, and was cut off mid-sentence. It now comes from the written analysis.
Ruled out explanations are evidence-backedThe ruled out section came from a separate model call with no evidence attached and sometimes invented confirmations. An explanation with no supporting data is now labelled unsupported rather than ruled out.
Chats survive a hard reloadA refresh landing while the server was restarting blanked the conversation list. A failed fetch is treated as "try again" rather than "you have no chats".
The first working product: the pipeline that turns connected tool data into one searchable knowledge graph, and the investigator on top of it.
New9
Ask why, and get a cited answerA diagnostic question runs a full investigation across your connected tools: narrowing to the relevant topics, pulling in related records, weighing evidence, arguing competing explanations and writing a cited conclusion, with each stage streaming live.
One knowledge graphRecords from every tool become typed entities and relationships. The same customer or version mentioned in different tools resolves to a single thing, and every extracted fact is checked against its source sentence before entering the graph.
Automatic deduplicationA four pass cascade collapses records that say the same thing even when worded differently. Every duplicate stays preserved and traceable to its source.
Topics and umbrella themesRecords that mean the same thing cluster into topics, and related topics roll up into named umbrellas, so you can enter your data by subject.
Answers carry real numbersInvestigations quantify impact directly from your analytics, payment and error data: total events, distinct users affected, success versus failure counts and a per version comparison.
Visual graph explorerSearch and filter by tool, node type and connection type, with colour-coded nodes sized by connectivity and a detail panel showing neighbours, verified facts and the raw source payload.
Timing, reference and similarity linksThe graph connects records that happened close together, records whose text names another record ticket or commit, and records that read as highly similar, giving you the deploy to error spike to failed charge chain.
See one record travel the pipelineWatch a real record journey through ingest, dedup, embedding, routing, extraction, verification, entity resolution, graph write and topic assignment.
Data shape alertsA watchdog compares each tool current data against its baseline and raises alerts when fields are added, removed or change type.
Improved2
Structured records carry real meaningRecords that are mostly fields with almost no prose, such as charges, analytics events and files, are described as a readable sentence built from their own field values, so they can be matched semantically instead of being invisible to search.
Fast lookupsShow me and how many questions return matching records and a count instead of sitting on the slow investigation path.
Fixed3
Citations open the real recordCitation chips carried only a reference, so the detail panel came up empty.
Topics no longer float disconnectedTopics survived the graph node budget but their member records did not, so topics rendered with no connections.
Builds recover from a crashIf a build crashed, the button refused to start a new one forever.
Releases before June 2026 predate this changelog. Want the detail behind a change? The documentation covers how each capability works.
