The crawler holds large content removals for your approval and shows per-run details, Studio gains a History page and keeps edge functions and forms in your branch, and QuantCode adds reasoning controls for self-hosted models and stops stuck turns.
Organisation owners and administrators can now register remote MCP servers on the new MCP Servers page under AI Services. A server must be reachable over HTTPS, and it can use no authentication, a bearer token or OAuth client credentials. Test connection checks that the platform can reach the server and lists its tools, and an optional allowlist limits which tools are exposed. Registered servers sync to QuantCode automatically. QuantCode connects through the Quant gateway, so the server's credentials are never sent to QuantCode.
When a full crawl no longer finds pages that were in the previous crawl, the crawler now re-checks each missing page before it removes anything. Small removals apply automatically. A large removal is held instead, and the organisation's owners get an email summary. A removal is large when it would remove more than 10% of the pages in the previous crawl or more than 1,000 pages. A new Content Removals screen on the crawler configs page lists held removals and the pages each would remove, and lets you approve or reject them. Pages in a held removal stay published until you approve. Approving removes those pages from the static site and from the AI search index. The feature is on by default for all crawl configurations, including existing ones. You can switch it off per configuration with the "Remove content deleted from the site" setting. With the setting off, the crawler does not remove missing pages.
New Crawler
Per-run details
Crawler runs with run tracking turned on now have a View Details action. It shows discovered, crawled, processed and error counts, and a breakdown of errors by type. It also has a paginated list of every page in the run. For each page, the list shows its status code, redirect target, error text, skip reason and the pages that linked to it, so you can trace a broken link to its source. You can filter the list by 2xx, 3xx, 4xx, 5xx, skipped or errored pages.
New Search
Field type overrides
A new Field Types card on the search extractors page lets you override how a field is indexed. For example, you can index a Unix timestamp field as int64 so that date-range filtering works. Changing a field type rebuilds the search index, and search stays available during the rebuild. Field type overrides are available only on QuantGov projects.
New Sentinel
Blind vulnerability confirmation and per-host rate limits
Sentinel can now confirm blind vulnerabilities through out-of-band callbacks, including blind SSRF, XML external entities (XXE), server-side template injection (SSTI) and blind or stored XSS. Start the optional callback service with the oob profile of the Docker Compose file, or set the oob block in the YAML configuration to use a callback server on your own domain. Without a callback server, Sentinel reports these classes as potential findings. --rate-limit sets the maximum requests per second that Sentinel's HTTP and GraphQL tools send to each host during an application scan. The default is 150, and 0 turns the limit off.
New Studio
History page
A new History section in the Studio navigation shows the active branch's commit history with a per-file diff for each commit, so you can see exactly what changed, when, and by whom, without leaving Studio.
New Studio
Edge functions and forms as code
Edge function and form definitions now live in your branch with your content. Each time you publish, they sync to the environment's CDN project. New environments get their functions and forms on first deploy, and promoting to production only needs a merge. Deleting an edge function in Studio now also removes it from the current environment's CDN project, and deleting a form turns it off there. Publishing only adds and updates definitions, so a deletion does not carry over to other environments through a merge. Form submissions stay with each environment and are never committed to the repository.
The Cloud Apps and AI Services usage cards on the subscription page show an error and a Retry button when usage cannot load.
Improved AI platform
Agent context management
The AI agent editor's Model tab now exposes auto-compaction settings. You can choose whether older messages are summarised once a conversation grows large, and the token count at which this starts. You can also set how many recent messages are always kept in full, and how many opening messages, such as the original request, are kept. When you customise a Quant-provided agent for your organisation, a new long context setting applies on the models that support it. You can use the platform default, force the 1M-token context window on, or force the standard 200K-token context window.
Improved AI platform
Slack bot setup
Setting up a Slack bot with your own app now provides a ready-made app manifest, pre-filled with the webhook URL and bot name. You paste one block into Slack instead of configuring events, scopes and App Home settings by hand. You can edit bot credentials after setup.
Improved Dashboard
Faster page loads
The environment details page now loads in about a third of the time it took before. The usage, invoice and payment cards on the subscription page are cached, so repeat visits load in well under a second instead of several seconds. The Studio search page appears at once while its metrics load in the background, and a progress bar and loading indicators show across the dashboard while an action is in progress. The cloud applications list is now cached. Changes you make in the dashboard show at once, but changes made outside it, through the CLI or a GitHub Action, can take up to a minute to appear.
Improved Forms
Export submissions through the Content API
The Content API's /form-submissions endpoint is now documented, so you can export a form's submissions programmatically instead of opening them one at a time in the dashboard. Pass form_fields=*, or a comma-separated list of field names, to include submitted values; without it each result contains only the submission ID, date and file count. Results are paginated at up to 500 per page.
July's Resources page now creates managed caches as well as object storage. Choose a capacity between 1 and 50 GB, then attach it to applications, detach it or delete it, as for object storage. Template deploys that include a cache provision it as an organisation resource, and caches created through an application are shown read-only with an App-managed badge. Enabling an application's cache now asks for confirmation and states the billing impact, and disabling one names the environments it redeploys.
Deleting a secret or environment variable that the currently deployed revision still references now warns you at the moment of deletion, and a persistent banner on the environment page reminds you until the next deployment.
Improved QuantCode
Self-hosted models
Building on July's local model support, QuantCode on QuantCDN can now use OpenAI-compatible models served from another machine on your private network or VPN, not only from the local machine. On QuantGov, self-hosted models remain limited to the local machine. You can now set reasoning effort for self-hosted Qwen3.8 and DeepSeek V4 models. Self-hosted Qwen3.8 now defaults to low effort, which roughly halved the turns and time per task in testing. QuantCode also now tells the model which input types it accepts, so a vision-capable model reads images directly. For every provider, not only self-hosted models, the exploration subagent now runs on a smaller, faster model where one is available.
Improved QuantCode
Stuck turns are stopped
QuantCode now stops a turn that starts repeating the same block of output, in its reply or in its reasoning. The turn no longer runs on until it reaches its output limit or you interrupt it. The message shows the text that was repeated. Lines that repeat normally in code, such as closing braces and imports, do not trigger it.
Improved QuantCode
Images from MCP tools are resized
Building on May's resizing of pasted images, images returned by MCP tools, such as screenshots, are now resized to the same limit before they enter the conversation. A 4K screenshot now uses about a sixth of the context it did, and the saving repeats on every later turn because the image stays in the session history.
Improved QuantCode
MacOS Terminal colour warning
The stock macOS Terminal does not support 24-bit colour, so QuantCode's interface shows there as dim grey text. QuantCode now shows a warning at start-up in macOS Terminal that recommends a terminal with 24-bit colour support, such as iTerm2, Ghostty, WezTerm or Kitty.
Improved Search
AI Search costs and spend caps
AI Search now appears in the per-interface spend caps introduced in June, so you can cap it on its own rather than only through the organisation-wide cap. The AI Services usage card on the subscription page shows AI Search as its own line, with page updates and AI answers listed separately. On projects where AI Search is enabled, crawler run rows show the AI Search cost of each run. Enabling AI Search now shows its usage-based pricing and asks you to acknowledge it before the feature turns on.
Sentinel 2.0 marks a finding Confirmed only when evidence captured from your target proves the exploit. A finding with supporting detail but no proven reproduction is Claimed, and a theoretical or disproven finding is Unverified. Reports link the captured evidence for each finding and add a Remediation section for each vulnerability class. Sentinel now reuses the sessions and credentials that it obtains during exploitation, so a report can show a privilege-escalation chain of several steps. Scans add a file upload track, a GraphQL track (set --graphql-endpoint), checks of client-side defences such as Content Security Policy and clickjacking protection, and WebSocket testing.
Pull the Sentinel image again to upgrade to version 2.0.
Improved Studio
In-browser deployments on the Deployments page
February's Deployments view listed only GitHub Actions workflow runs, so it looked empty for Quant Builder mode projects that publish from the browser. In-browser deployments are now recorded too — who ran each one, what it published, page counts, timing and status — and the whole organisation can see them. The Deployments toolbar also gains a Preview button that opens the environment's preview URL, and a chip next to it shows how long ago the current branch was last published from the browser.
Improved Studio
Branch status and publish prompts
Each environment branch now shows how many commits it is behind the production branch, or that it is up to date. An update action named after the production branch (for example, Update from main) merges in the latest production commits, reports any conflict, and then prompts you to publish the environment. The change-count banner now names both branches it compares, for example "12 changes on dev not yet in main". Entry badges now read Changed and In sync instead of Modified and Published, with a tooltip that names the production branch. In Quant Builder mode, a merged pull request keeps its Build & Publish prompt when you return to it. The branch selector now also appears on the History, Forms and Search toolbars.
Improved Studio
Site search per environment
Site search now provisions a separate index for each environment. The widget reads its site id from the environment at build time, not from a value committed to the repository, so merging a development branch into production no longer carries the wrong index with it. Purge, re-index and boost actions on an environment URL act on that environment's index. The indexing counter reports pages sent for indexing and polls until the index reflects them.
If your search results already list the same page twice, run a purge and then a re-index once.
Improved Studio
Site search filters and facets
Studio site search, introduced in July, now indexes each entry's frontmatter as metadata when you publish, so you can build faceted or pre-filtered search pages. A new Filters tab in the search settings lets you define the fields to filter on, with a type and a label for each. You also choose which fields show as filter controls in the search widget.
Republish pages that you published before August so that they get their metadata.
Improved Studio
AI assistant change lists and image requests
When an assistant turn changes files, a list of every file it changed is now added after its summary. The list comes from the edits actually made, so you can review the full set of changes from the chat. When a design needs an image that the assistant cannot create in code, such as a logo or a photo, it now asks you to drop the image into the chat. It places the image in the same turn.
Proxied responses with a fixed length, such as large files, reached visitors only after the whole response had downloaded from the origin. They now stream as they arrive, as server-sent events and chunked responses already did. Separately, a request to an origin that stopped responding could wait longer than the configured origin timeout. Requests now end at the configured timeout.
Fixed Crawler
Run status reflects the outcome
Runs that finished cleanly could show as Stopped rather than Completed. The run list now shows whether each run completed, failed or was cancelled, and old cancelled and failed runs are now removed on the same schedule as completed ones. On older crawl configurations, the run tracking toggle could show as on while crawls ran without tracking, so run progress stayed on "Starting..."; the toggle now shows the tracking setting that the crawl actually uses.
Fixed Crawler
Non-recursive sitemap entries
Setting recursive: false on a crawler's sitemap entry, through the API or Terraform, had no effect, so the crawler still followed the sitemap index references in that sitemap. It now loads only that sitemap's own URLs.
Fixed Dashboard
Uploaded HTML files serve as pages
HTML files added through bulk upload or ZIP archive upload were published as files rather than pages, and uploaded index.html files were downloaded by browsers instead of displayed. HTML uploads are now published as pages and served as HTML.
Fixed Quant Cloud
Container forms and resource detach
The Container Dependencies add and remove buttons on the create-environment form did not work; you can now add and remove dependencies there. On the container edit page, removing an environment variable, secret, mount point or exposed port removed the first row rather than the one you clicked; the correct row is now removed. On the Resources page, a failed detach now shows the reason instead of a generic error, and the detach confirmation warns that the environment redeploys immediately and may be briefly interrupted.
Fixed QuantCode
Replies and message order in existing sessions
From mid-August, a new message in an existing session could get no reply or disappear from the transcript, with no error shown. Returning to a session could also hide its oldest messages. Sessions now reply normally, messages stay in the order you sent them, and a message that fails to send is reported instead of silently cleared.
Fixed QuantCode
Compaction keeps the current task
When a long session was compacted mid-task, the agent could lose track of what it was working on and report that there was no request to act on. Compaction now carries the in-progress task through.
Fixed QuantCode
Interrupted turns no longer leave a session busy
If QuantCode quit, crashed or the computer went to sleep part-way through a turn, the session could show as still working indefinitely, and follow-up messages queued in it were never sent. Interrupted turns are now closed when the session is next opened, so the progress indicators clear and queued messages are delivered.
Fixed QuantCode
Model selection is kept
A model chosen with /model in a newly opened session could revert to the session's last-used model, and a per-agent model choice was lost when QuantCode restarted. Your choice now stays in effect and persists across restarts.
Fixed Rules
Matching on campaign parameters
Rules that match on utm_* campaign parameters, such as utm_source, utm_medium or utm_campaign, did not fire, even though they saved without error. These rules now fire, and URLs with different campaign values are now cached separately. Click identifiers such as fbclid and gclid are still ignored for caching and cannot be used in rule conditions.
Fixed Search
Extractor test dialogue
An empty dialogue could appear over the search extractors page after an interaction and block every click. This is now resolved.
Fixed Studio
Creating and editing entries
Creating a new content collection entry could open a blank preview. Collections with a required object field also produced entries that failed validation. New entries now open in the preview and include their required object fields. When you create an entry, you can now choose its collection. The automatic slug no longer overwrites a slug you are typing, and Duplicate opens the new copy. On multilingual projects, AI assistant edits could go to a file that the site does not render, so Preview did not change. The assistant now follows the project's locale settings and reports the file it wrote.
Fixed Studio
Environment pages show their own environment
On a non-production environment, the Forms page, edge functions list and edge function previews showed the production environment's items, not their own. Each environment's pages now show that environment's forms, edge functions and preview domain.
Fixed Studio
Re-imported environments keep their settings
Re-importing a deleted Studio environment could reset its per-environment settings, such as its site search setup. Re-import now keeps the environment's own settings as they were.
Fixed Studio
Edge function islands render
An edge function island is a placeholder in an entry that shows fallback text until the browser loads its content from one of your edge functions. Edge function islands inserted from the toolbar showed their fallback text indefinitely. They now load. The island picker also adds options to load on a media query, poll, and set the request method. Islands need the project's inline functions option, so Studio now turns it on when you first deploy an edge function or when functions sync to an environment.
Purge or republish cached pages that contain edge function islands.
Fixed Studio
Files missing from the file tree
Files present in the repository could go missing from Studio's file tree, so builds could run without them and produce error pages. The Functions tab could also appear empty while your functions were still in the repository. Missing files are now fetched again on the next sync, and the Functions tab shows every registered function.
Fixed Studio
Frontmatter dates and lists
Frontmatter timestamps with a numeric UTC offset, such as 2026-08-24T19:57:45+00:00, were not recognised as dates in the preview and in-browser builds; they are now read as dates. Dates in the preview could also differ from the published site, depending on your time zone; the preview now shows the same dates as the published site. Lists of objects written with the dash on its own line collapsed into a single item; every item is now kept.
Fixed Studio
Site migration crawl step
Migrating a site into Studio could fail at the crawl step. This is now resolved.