New updates and improvements about V7 Go

A short one, weighted towards the API.

API

• ID mappings in template exports: A project template export can now include optional ID mappings, which makes moving a template between projects a mapping exercise rather than a manual rebuild.

• Tools keyed by name: Discoverable V7 Go tools are returned as a map keyed by name, so a client can look one up without iterating the array.

• Capability-based gateway authorisation: The gateway's data_store_only_api claim has been replaced with run capabilities, a more granular model. Integrations relying on the old claim need to move across.

Interface

• Citations are underlined for everyone: The underline styling for citations has left its flag and is on by default.

• Workspaces without an organization: Workspaces that do not belong to an organization are listed consistently across pages instead of appearing only in some views.

• Raw and formatted debug views: The sandbox debug page and the shared trace view can toggle between formatted and raw data.

Three things worth knowing about if you build on V7 Go rather than only work in it.

Building

• JSON Schema draft 2020-12: JSON properties accept draft 2020-12, so schemas using the modern keywords no longer have to be rewritten down to an older draft.

• Webhooks renew themselves: Native webhook subscriptions are renewed automatically instead of expiring and needing to be recreated.

• File requests carry a reason: A file request now requires a description and a reason, which show up wherever the request is reviewed.

Accessibility

• Project tables by keyboard: Table actions are reachable and operable from the keyboard, and the underlying table semantics have been corrected for screen readers.

• Tab leaves the code editor: Tab now moves focus out of a code editor instead of being captured by it.

• Onboarding motion: The onboarding flow no longer animates on its own, and the pause button it needed has gone with it.

Administration

• Workspace access over the API: A new organization endpoint manages workspace access permissions programmatically.

• Plan usage counts consistently: Usage reporting and limit enforcement now count the same way, so the number you see matches the number that gates you.

Trash picks up where the archived-runs view left off, and a few things get shorter names.

Workflows and connections

• A Workflow runs tab: Building on the archived-runs view, Trash now has a dedicated Workflow runs tab and a page per workflow, so a deleted run can be found and recovered without going back through the workflow itself.

• Connection visibility: The owner of a native connection can now set who is able to see and use it.

• Skill names lose the prefix: Agent Skill names no longer carry the internal workbook- prefix.

Reading and tables

• Lighter assistant prose: Assistant messages in chat have lighter type and more deliberate spacing.

• Column group headers: A group header now spans the columns it covers, so a wide table's structure is readable at a glance.

• Sidebar renders open folders only: Large project trees scroll without the lag that came from rendering every collapsed branch.

• Faster KV search: JSONPath search scans the key prefix range directly rather than walking the whole space.

Access

• Manage access replaces the invite dialog: Workspace invitations now go through the Manage access dialog.

• Seat denial clears itself: If you gain access by another route, such as a group membership, the seat denial no longer persists.

• Focus stays put: A hover card or popover opened by keyboard focus leaves focus on the trigger.

Removals

• question_request_id is gone: The deprecated field has been removed from the add-query request; clients still sending it need to stop.

• HubSpot is not pre-installed: New workspaces no longer come with the HubSpot integration already added.

Mostly a week about not losing things: messages that used to disappear before you had read them now wait.

Notifications

• Validation popovers stay open: A validation popover now stays until you close it, so a message cannot vanish while you are still reading the field it refers to.

• Essential toasts stay open: Toasts that carry something you need to act on no longer auto-dismiss.

Mission Control

• Highlight a series from its legend: Hover or click a legend item to highlight that series in the chart.

• Line chart points by keyboard: Chart points are reachable by keyboard and exposed to screen readers.

Errors and indexing

• Temporary versus persistent: A new temporary_error label separates errors worth retrying from ones that will not resolve on their own; the old internal tool error label has been renamed to match.

• Context index capacity: Identity queue limits have been raised, so large indexing batches no longer throttle.

Interface

• Settings and Trash match: Workspace settings pages and Trash now follow the same design as the rest of settings.

• Old question cards removed: The inline question card no longer appears in older chats.

• Recommended badge: Now uses the theme's warning colours rather than its own.

Under the hood

• OCR geometry is validated first: Invalid geometry and spans are rejected before positioning data is built, so a bad page fails cleanly instead of producing wrong coordinates.

• More API key receivers: HTTP API key authentication accepts additional receiver methods.

A batch of MCP work this week, plus a set of changes that make the interface calmer and easier to read on a small screen.

MCP

• Collections as workflow input: run_workflow now accepts a collection directly, so a batch run no longer needs a preprocessing step to flatten rows first.

• Approval cards say what the tool is for: An MCP approval card now shows the purpose of the request, so you can judge it without opening the tool definition.

• The MCP section opens by default: Integration details are visible as soon as you land, instead of behind a disclosure.

• Go Agent V2 knows the catalogue: V7 Go MCP tool names and descriptions are in the Go Agent V2 system prompt, so the agent can tell you what is available rather than guessing.

Breaking API change

• run_mcp_tool argument rename: The arguments_json parameter is now arguments. Calls using the old name need updating.

Reading and accessibility

• Automatic motion can be turned off: A new setting stops interface elements animating on their own.

• Narrow screens reflow properly: Layouts now reflow correctly at 320 and 768 CSS pixels.

• Docked chat is opaque: The docked chat cards and the composer no longer let the page show through.

• Field types show as icons: The run Input card marks each field's type with an icon instead of a text label.

• Longer feedback fits: Plan adjustments grow to fit their feedback text instead of truncating it.

Clearer states

• Empty workflow chats explain themselves: A chat with no rows in a workflow shows a message instead of an empty table.

• Archived run restoration: The messages about what can and cannot be restored have been rewritten.

• Diagram checks: A diagram check stays gray until it actually completes.

Performance and security

• Runs created in one batch: A workflow run and its collection rows are written together, which cuts database load and removes a class of partial-write inconsistency.

• Reporting jobs are workspace-scoped: Each reporting job is now authorised from its own workspace row, closing a cross-workspace path.

• brotli CVE-2026-111497: Patched in both the library and the production image.

Voice typing is out from behind its feature flag. It is on for everyone, with nothing to request and nothing to switch on.

If you would rather talk than type, you can now dictate instead of typing a long instruction into a chat. Everything else about the conversation works the way it did.

Availability

Voice typing is available to all users. The flag that limited it has been removed.

Generated decks survive the cases that used to break them, a connection the provider has revoked stays revoked, and a retired connector can finally be disconnected.

Slides

• Decks that render: A generated deck survives a text run with no font name, a slide field the generator left out, and an image cropped past its own edge.

• Open a deck without the chat: A slide deck output opens without needing access to the run chat that produced it.

Connectors

• A revoked connection stays revoked: A native connection the provider has revoked is no longer refreshed behind your back.

• Disconnect a retired connector: An owner can disconnect a Slack connector that has been retired, rather than leaving it in place permanently.

Tables and views

• Unsaved columns stay out of saved views: View updates route through one function that strips unsaved columns, so configuring a view no longer writes half-made state into it.

• Model lists show what you use: Report provider lists drop models nobody in the workspace runs.

Keyboard and pointer

• Citations are reachable: An underlined claim can be reached and opened with the keyboard.

• Tooltips behave: A tooltip can be hovered and dismissed with Escape.

Security and housekeeping

• Refreshed dependencies: Certificate handling and the Python tool lock were updated to drop exploitable library versions.

• Retired tools removed: The superseded PDF and slide-deck harness tools are gone, along with a chat listing parameter that no longer applies.

• Steadier under load: Logging moved onto a non-blocking handler, lock manager mailboxes moved off heap, and a vector index is left alone when it is already valid.

Archived workflow runs have somewhere to live and a view to read them in, report tables show when things were created, and an invitation to a full workspace no longer just fails.

Archived runs

• A view of what was archived: Workflows have an archived-runs view, and archived runs are grouped in the workspace Trash rather than mixed in with live work.

Reporting

• Creation dates in report tables: A report table shows when each thing was created, so age reads alongside usage.

• Retired coverage routes: The reporting coverage routes that read job records are gone, which is a breaking change for anything calling them directly.

Seats

• An invitation at the seat limit: Someone can be invited to a workspace that is already at its seat limit, rather than the invitation failing outright, and the seat-denied email now states the pool and the limit it ran into.

Connectors

• Only the scopes that were registered: An MCP connection is granted the registered overlap of the scopes it requested, rather than everything it asked for.

Chat and files

• One file, once: A generated file appears a single time in the chat response.

• Messages stay visible from a PDF: Selecting a message from the PDF preview keeps it in view, and All data keeps Download all through the redesign.

• Ask Go keeps to itself: The Ask Go side chat no longer leaks into the project page state.

• Tighter reasoning: Expanded activity uses its vertical space better.

Speed

• Faster research lookups: Source research lookups are quicker and easier to use.

• A lighter project page: The project API response is cached under one shared query key and table rows are looked up by id, which shows most on a large table.

Claude Haiku 5.5 is available, the chat data pane is rebuilt around resource tabs that hold their place in the URL, and a report keeps the cost of a chat after somebody deletes it.

Models

• Claude Haiku 5.5: Available, and automatic model selection has been updated to take it into account.

• Cheaper summaries: Summary work prefers a lighter model where the job does not need a frontier one.

The chat data pane

• Resource tabs that survive a reload: The data pane and the sidebar overlay both use resource tabs, and the tabs stay in step with the URL, so a particular view of a chat can be linked to and come back the way you left it.

• Uploads and cited sources are separated: All data tells the files you attached apart from the sources a run cited, which are two different questions, and deduplicates by file ID so one file appears once.

• Work runs fold away: A run folds into a single activity row and file cards render in the new style, so a long chat reads as a sequence rather than a wall.

• Tabs survive a full session store: Open chat tabs are kept when session storage fills up, instead of disappearing.

Reporting

• A deleted chat keeps its cost: Chat cost from a deleted chat still counts in the Users and Chats reports, deleted chats appear in the Chats tab, and token report rows carry a deleted marker, so a total does not quietly shrink when somebody tidies up.

• Access rechecked on a workspace switch: Moving between workspaces rechecks Reports access rather than carrying the previous answer over.

Runs and retention

• Retention settings: A workspace carries retention settings for workflow runs, and deleting a run routes through the archival path, so what is kept is a decision rather than an accident.

• A workspace-level workflow list: Every workflow in the workspace in one list, with a workspaces tab on the user record.

Naming, connectors and fixes

• Workflow, not project: Interface labels say workflow where they used to say project, matching the rest of the product.

• Sentry in the default connections: The Sentry connector is included by default.

• Smaller irritations: The file viewer stays open when Escape closes a menu, PDF hover previews resolve without OCR, and dialogs carry a name and a complete keyboard contract.

• Safer slide scenes: SVG scenes are sanitised and part IDs isolated.

Table and toolbar actions fire on the first click, hover-only actions stay visible when you reach them with the keyboard, and a connector that refuses a request now says what actually went wrong.

One click, and the keyboard too

• Reset, Recompute and Skip act on click: The table cell actions run when you click them instead of needing a second go.

• Toolbar commands act on click: Rich text and slide text toolbar commands activate on click.

• Hover-only actions survive the keyboard: An action that only appears on hover stays visible when it takes keyboard focus, so it can be reached without a mouse.

• Tab leaves the table: Tab moves out of the project table rather than cycling inside it, and a focused cell is scrolled into view rather than left off screen.

Connectors that explain themselves

• A refusal reads as a refusal: A 403 from a connection that is already authenticated is reported as what it is, so the problem reads as a permission rather than a broken connector.

• No phantom administrator: The misleading notice telling you an administrator has to enable something is gone from OAuth connectors that need no such thing.

• Switch client from the connection widget: The in-chat connection widget can move a connector onto the V7 template client, so the switch no longer means leaving the conversation for the settings page.

Reporting exports

• A line per API key: The Users CSV writes one line per API key rather than folding them into a single row, so machine traffic can be read key by key in a spreadsheet.

• One canonical email column: The Users and Chats CSVs both name the email column user_email, so two exports can be joined without renaming a column first.

Slides

• Image crops survive the export: A cropped image renders in the unit PowerPoint expects, so a deck looks the way it did in V7 Go.

• Fonts load: Google Fonts load reliably in slides rather than falling back part of the time.

Chat

• What it says to you stays readable: Text meant for you stays prominent while work is running underneath, instead of being buried in the activity around it.

• Steering continues cleanly: A conversation redirected mid-run picks up from the steer rather than stumbling over it.

Underneath

• Archived runs stay archived: A calculation cannot start on an archived run, so an archive stays a record rather than becoming live work again.

• Fewer wasted cache reads: A cache miss fills a placeholder instead of reading twice.

A workspace can now set where its inference is allowed to run, and the policy is enforced on the way to every provider. Azure OpenAI covers the current GPT models, and OpenAI calls route through EU or US endpoints.

Where a model is allowed to run

• A policy the workspace owns: A workspace holds its own inference policy, and a request that falls outside it is refused with a reason rather than quietly sent somewhere else.

• The same rule whichever provider serves Claude: Region policy is enforced across Anthropic, Vertex and Bedrock, so a Claude run honours the region it is held to no matter which of them answers.

• Gemini, Vertex and Azure OpenAI as well: The Gemini, Vertex and Azure OpenAI execution paths each check the policy before they run.

• Embeddings are covered too: Embedding requests follow the same policy, so a document indexed for search cannot leave a region that a completion would not have left.

Providers

• Current GPT models on Azure OpenAI: Azure OpenAI coverage is brought up to the current GPT line, so a firm that buys its OpenAI capacity through Azure is not a generation behind.

• The Azure OpenAI Responses transport: Supported, so calls use the transport Azure expects.

• OpenAI through EU and US endpoints: OpenAI inference is routed to a regional endpoint rather than a single global one.

• Background Responses for workflow calls: OpenAI calls in a workflow use the background Responses path, which holds up better on long work.

Underneath

• Entities insert in batches: Added entities and their fields go in as multi-row batches, which shows most when a run creates many rows at once.

• A usage-write error stops taking the job with it: A failed usage write no longer fails the scheduled job it was part of.

• Steadier locking across nodes: A node-local lock manager replaces the previous global lock, removing a source of contention under load.

MCP, Skills and Code now sit in one Code tab with a search over the connector catalogue. Outlook can be pointed at a specific mail folder, including a shared mailbox, and user emails are stripped from error reporting.

One place for what V7 Go connects to

• MCP, Skills and Code in one tab: Integrations and Custom are merged into a single Code tab holding MCP, Skills and Code, so a connector, a skill and a snippet are no longer three places.

• Search the catalogue: The connector catalogue has a search under the banner, which matters more the longer the catalogue gets.

• Setup appears where you are: The in-chat connection setup widget shows when it should, and the guide for bringing your own client credentials gained the step it was missing.

Outlook

• Pick the mail folder: A native Outlook connection can be pointed at a specific mail folder, with its options and errors shown rather than guessed at.

• Shared mailboxes: Connections store their OAuth scopes and request shared-mail read access, so a shared mailbox is reachable rather than silently out of scope.

Properties

• A new provider for Select properties: TypeSafe is available as a provider, with Jev selectable for Select properties, so a classification step has another option behind it.

Organization

• Seat summary at a glance: Billable seat summary cards show where the seats are going without reading the roster.

• Initials when there is no logo: An organization without a logo shows its initials rather than a blank.

Privacy and reliability

• User emails stay out of error reporting: User email addresses are anonymised and URL secrets stripped before anything reaches error and analytics reporting.

• Secrets are hashed in cache keys: A secret used in a cache key is hashed rather than stored in the clear.

• Facets retry on a manual recompute: A failed file facet is retried when you recompute, so one failure does not leave a gap you have to find yourself.

Accessibility

• Errors are announced and focused: A text field error is linked to its field, and submitting an invalid form moves focus to the first field that needs attention.

Workflow memory is now called Workflow Guidance. CARTO joins the connector catalogue on your own OAuth credentials, and an edit in a table survives the row reloading underneath it.

A clearer name

• Workflow memory is Workflow Guidance: The name says what it is: the standing guidance a workflow follows, rather than something that remembers on its own.

Connectors

• CARTO, on your own credentials: CARTO is in the catalogue and connects with your own tenant OAuth credentials, so the connection belongs to your account rather than to a shared client.

Edits that survive a reload

• Reference picks are kept: An unsaved reference selection survives the row refetching, so a pick is not lost to a refresh you did not ask for.

• Cell edits stop going missing: An edit in a cell is kept rather than dropped.

• Uploads that start early still finish: A library upload begun before the library has loaded, or during a list reload, completes instead of being abandoned.

• Fewer requests behind a table: The imports list is fetched once rather than for every cell on screen, which shows most on a wide table.

Reporting

• A keyboard date pick holds: Picking a date with the keyboard no longer resets the report period.

• Field calculations on the user overview: Field calculations appear among a user’s actions, so their work is counted where it is read.

Underneath

• Steadier coordination across the cluster: Cluster locking moved onto Redis and debouncing onto an atomic cache claim, which removes a class of stall under load.

• More room for outbound traffic: Outbound HTTP pools are sized for their traffic, and a pool checkout timeout comes back as an error rather than as a hang.

A Workflow file mentioned in a chat is now a link that opens in the chat’s own file preview. Chats are named from what you asked rather than from the workflow, and GPT-6.1 Sol and Sonnet 5.5 are selectable.

Files where the conversation is

• Workflow files are links: A reference to a Workflow file renders as a link in chat rather than as text you have to go and find.

• They open in the preview: Following one opens the file in the chat file preview, so a document can be checked in place rather than in another tab.

Chat

• Named from what you asked: A chat takes its name from your request rather than from the workflow behind it, so a list of chats reads as a list of questions.

• Home opens a full-page chat: A chat started from Home opens as a full page instead of a panel.

• Search and filter in the sidebar: The sidebar list header carries search, a create menu and a chat filter, so finding an earlier conversation is a search rather than a scroll.

• No duplicate of your own message: The pending copy of a sent message disappears once the server copy arrives.

Models

• GPT-6.1 Sol and Sonnet 5.5: Both are selectable in chat.

Connectors

• A Google Analytics template: Google Analytics is available as an MCP template, so connecting it starts from a prepared setup.

• Longer-lived connections: MCP OAuth discovery requests the offline_access scope where the provider supports it, so a connection lasts rather than needing reauthorising.

Editing without losing work

• A truncated value cannot be half-edited: A field showing a truncated value stays read-only until the whole value has loaded, so an edit cannot overwrite the part you could not see.

• The newest value wins: Concurrent saves of the same field are coordinated so the latest value is the one kept.

Reports and interface

• A date range that sticks: The committed report date range survives a URL change rather than resetting underneath you.

• A failed page is not asked for again: A report page that failed to load is not re-requested in a loop.

• Settings and logos: Settings layouts are aligned and logo uploads are limited to formats that work.

• Named dialogs: Every dialog takes its accessible name from its own header, and the accessibility scan covers landmark and heading rules.

Reports carry model usage per user and export it to CSV, a field calculation counts toward activity, and a workspace at its seat limit refuses the seat rather than the login.

Models

• Full Opus reasoning levels: The complete range of Opus reasoning levels is exposed, so how hard the model thinks is a setting rather than a default.

Reporting

• Model usage per user: A per-user model usage series is available and exports to CSV, so a model’s cost can be read against the people using it.

• Field calculations count as activity: A field calculation counts toward user activity in the Users and Overview views, so someone whose work is calculations rather than chats appears as active.

• Per-user activity in organization reports: An organization report breaks activity down by person, so adoption can be read across workspaces rather than only in total.

• Reports assemble in parallel: Independent reporting steps run concurrently, so a report waits on the slowest step rather than the sum of all of them.

Seats

• A full workspace refuses the seat, not the login: When a workspace is at its seat limit, a first login is denied a seat without the login itself failing, so the person gets a clear outcome rather than a broken sign-in.

• Seat distribution in the roster: Billable seat distribution is returned alongside organization users, so where the seats went is answerable.

Chat

• One slow chat stops blocking the rest: A chat with messages pending no longer holds up the others, so a long run in one place does not freeze work somewhere else.

A report now says where its data starts instead of drawing zeros over a period it cannot cover. GPT-6 Luna, GPT-6 Sol and Opus 5.5 are available, and Hub files can be moved in bulk.

Reports say what they do not know

• Where the data starts: A report shows the point its data begins rather than plotting zeros before it, so a flat start is not mistaken for a quiet month.

• Incomplete is marked as incomplete: A range that reaches back before a workspace’s extracted history reads as incomplete, and the earliest extracted hour is recorded on each checkpoint.

• Partial data stops being permanent: Spend in a long-deleted chat no longer leaves a report reading Partial data forever.

• Freshness checks run in parallel: The freshness check runs its queries concurrently, so a report waits less before it appears.

Models

• GPT-6 Luna and GPT-6 Sol: Both are available.

• Opus 5.5: Available alongside them.

Longer chats, less lost

• Completed tool payloads are compacted in place: A finished tool payload is compacted within the turn it belongs to, so a long turn keeps working instead of running out of room.

• Recovery survives context falling away: When earlier exchanges leave the context window, the ability to recover the run survives with it, and compaction budgeting is corrected.

• Files it writes are checked: An Office or JSON file produced in a run carries validation warnings rather than arriving quietly broken.

• Reporting over the MCP bridge: Reporting can be queried through the MCP bridge, so a usage question can be answered in the conversation.

Hub files

• Bulk drag and drop: Several files and folders move together, by dragging or through a bulk move, instead of one at a time.

Reporting detail

• Cost share on property reports: A property report carries the workflow cost share, so a field can be read against the workflow it sits in.

• API key spend on the Users report: The Users report shows API key spend with attribution filters.

• Budgets read against regional settings: A reported budget is linked to the regional settings that define it.

• Filters stay where you put them: Filter positions survive an update, and a tooltip drops its footer when there is no breakdown to show.

Interface

• The connector catalogue keeps growing: Exa is in the catalogue, with its logo on the connector card.

• A model picker that fits: The chat model picker matches its intended size.

• Tidier chrome: The workspace logo lines up with the rail icons in the collapsed sidebar, sidebar labels use the space left by hidden action menus, and action menus carry a plain dots icon.

• Drafts stop bleeding between tabs: A case draft no longer echoes a stale value from another tab.