Branches in chat and one model for your chat appsPick the branch a chat works on right in the composer, and choose the model Slack, Teams, and Google Chat run on for your whole organization. Settings gets a search that finds any setting, connection, or teammate.
- Chat: Pick the chat’s branch in the composer, create one, or check out a merge request. One picker shows provider, model, and reasoning side by side, and after you connect a tool the agent suggests what to set up next.
- Chat apps: Org admins choose the model Slack, Teams, and Google Chat run on, for the whole organization and per channel, instead of each sender’s own pick.
- Resolve: Approval emails and cards say in plain words what is paused and why. The agent posts its triage reply even when it tags no one, and a triage chat shows one card instead of two.
- Settings: Search finds any setting, connection, or teammate, and the breadcrumb reaches memory, knowledge, and artifacts. Each user can hide the CloudThinker Assistant bubble.
- Billing: Free organizations can invite unlimited teammates.
- A slow sandbox setup no longer ends a review right after it starts, and quiet repositories no longer raise a staleness alert.
- Resume on a Code Audit opens the new run, and Fortress teammates keep going until three passes in a row find nothing new.
- Cost health checks save reliably again.
- A connection that breaks in the background shows a clear sentence on what went wrong.
- The Home timeline counts an incident as running only while its agent is still investigating.
One Inbox, a calmer Home, and a flatter designOne Inbox gathers everything that waits on you, and the homepage is calmer, with a timeline of recent work. The whole app moves to one flat, consistent design.
- Home and Inbox: The Inbox has a fixed icon in the left rail, with notification settings one click away. The homepage shows status chips, a timeline of recent work, and a right-hand rail for incidents, products, and automations.
- Resolve: Investigations show which incidents are still alerting, a worse alert on an open incident reaches triage, and alerts with one cause group into one family. Fix options sit right under the causal graph, each RCA finding shows the checks behind it, and the next RCA learns from a responder’s “wrong” mark.
- Review: Auto-Resolve fixes conflicts one hunk at a time, settles lockfile and whitespace conflicts first, and keeps its follow-up merge request when the code host fails. A Learning can become a reviewed pull request, Learnings that developers keep rejecting are skipped, and Review setup ends with an offer of a first review.
- Cyber: Code Audit library research reads the web and scans on demand. Pentest agents share existing credentials and discoveries, and team progress survives restarts.
- Chat: Answers link to app pages, and metric charts use the full width. A Room keeps its decks and pinned files beside the chat, projects and folders remember their repository and where chats run, and the CloudThinker Assistant guides you around each page.
- Optimize: The agent can close cost anomalies, you can retract them, and high ones send an alert. The Overview is grouped into cards, and the last cloud account you picked is remembered.
- Connections: Grafana reads about 25 more data sources, among them VictoriaLogs, VictoriaMetrics, Zabbix, Sentry, Jaeger, and ClickHouse. The agent can run any FPT Cloud Console operation, and any MCP client can connect to CloudThinker.
- Onboarding and Settings: New users pick their modules and get a six-step checklist. Settings is one dialog with account tabs, a breadcrumb, and a shareable URL.
- Bring your own BytePlus key, and your own Claude keys default to the newest models.
- Chat, triage, and RCA keep working in Light mode when credits run out.
- Dropdowns, spinners, and dialogs share one look across the app.
- Saved memory can be browsed by type.
- Finding titles in Optimize are shorter and easier to scan.
Codebase Atlas, one incident place, and chat rewindCodebase Atlas is generally available, and Loupe leaves Beta. Each incident now has one place in chat, and you can rewind a chat or edit a message and resend it.
- Review: Codebase Atlas opens on the last 7 days and loads in about half the time, and the Review sidebar groups its pages under four labels. Release risk tells your team whether a pushed tag can ship, and one GitHub organization can connect to several workspaces.
- Resolve: Updates for an incident go to one place in chat instead of several channels, and Investigation shows one clear next step. Incidents show a severity icon, the queue sorts by severity, newest, or recently updated, and channels show names instead of IDs.
- Cyber: Code Audit library monitoring always runs, push audits are on by default, and a burst of pushes runs one audit. Each repository has one Code Audit page, a library page can open the upgrade pull request, and every run saves a PDF report.
- Chat: Rewind a chat to an earlier reply, or edit a message and resend it. The sidebar is flatter, one ring by the composer shows how full the context is, and first replies start faster.
- Automations: One automation can run on several schedules, each edited inline without cron, or on its other triggers alone with no schedule. New automations start from templates for any connected tool.
- Connections: Connect Cloudflare with IP-restricted API tokens. A new Get apps and integrations button in the sidebar gathers the CLI, Slack, Microsoft Teams, and Google Chat.
- Workers: Admins can limit which hosts a Worker and a managed sandbox may reach. Both refuse to run rather than run open.
- Billing: Plans show the Free tier in the same layout as paid plans, and yearly plans show their monthly price. Owners start card-backed trials themselves, and Enterprise shows Custom workspace capacity.
- A failed action shows the real reason and one next step, and out of credits no longer looks like a permission error.
- Reviews recover on their own after a failure, a timeout, or a sandbox outage.
- Stopping or steering a reply keeps credit usage correct and releases unused holds, so a plan change is no longer blocked.
- Google Chat spaces can start listening on the On-call page, and admins can move the notification space again.
- Security fixes harden sign-in, SAML, SSO, and SCIM.
Faster, more careful reviews and chat foldersReview now runs on a new default engine that finishes in about 5 minutes and checks each finding against nearby code before it posts. Chats get folders, and long agent runs keep working while older history is summarized.
- Review: Every workspace that never picked an engine, and every Fast review, uses the new engine. The Reviews list opens on your own merge requests, the bot answers review and autofix, and Release risk analyzes new tags on its own.
- Chat: Folders hold chats, their shared files, and custom instructions for the agent. You can share a chat while the agent is still replying, and sending a message while an approval card waits declines it and runs your message.
- Workers: Run in shows in the chat + menu before your first Worker exists. Reconnect an old conversation on an offline machine from its chip, and the Worker settings dialog says why a machine lost access and how to bring it back.
- Cyber: Ultra Security Testing keeps going while it makes progress and stops when it stalls. Deep dive, the most advanced testing mode, is available, and exported findings get a cover, summary, and findings index.
- Connections: Connecting a provider needs only its credential, and saved credentials can be edited again. An expired GitLab token turns the connection red with a fix button, and PRTG Network Monitor is a new connection.
- Automations: An automation can send each run’s result to a Slack, Teams, or Google Chat channel. An admin can take over an automation whose owner left.
- Optimize: CloudThinker asks a person before it closes a finding and reads your team’s comments and edits first. A new account’s first scan uses this month’s spend so far.
- Memory and BYOK: Each remembered fact in memory names the chat and date it came from. A BYOK key can also run background tasks such as chat titles and recaps.
- Images you add while the agent is working reach the agent right away.
- Replies stay in the language of your latest message, also in long conversations.
- Picking one of your own Workers under Run in no longer breaks the home page.
- In Slack, Teams, and Google Chat, CloudThinker reads the alert you reply under before it answers.
- Failures no longer show raw server errors.
Check after this release. Only organization Owners and Admins can read audit logs now. Incident investigations can now wait for approval, so make sure your approvers watch the notification bell. Automations and webhooks wait only when their Approval mode asks a person. By default they use Auto · Rules + AI · Refuse, which refuses instead; to make them wait, change the Automations row under Run types.
- Approvals: Agents propose Rules, and Find rules drafts them for your workspace. Set approver groups with a two-person rule, give time-boxed grants, and dry-run Rules against past history.
- Chat: Side chat (
/btw) asks a quick side question in a fork next to a running chat. Shared chats show who is present and typing. Stop halts the running reply at once, and long chats load older messages about 3 times faster. - Memory: A nightly Dreaming run tidies workspace memory and custom skills. Workspace Admins can run Dreaming, review its work, and edit items in More → Workspace settings → Memory.
- Review: A review posts a few focused findings that stay stable between pushes. A push or deploy no longer stops a running review, and the Review page now opens on the revamped Overview.
- Cyber: Resume an interrupted test run. A cloud test run waits in the queue instead of failing.
- Optimize: Developers and above can edit findings and see each change in Activity. Workspace Admins get one monthly digest instead of an email after every scan.
- Resolve: A flapping alarm that a person closes as a false alarm stops looping. Root cause analysis reports say how to check each claim again.
- Connections: A failed connection explains the problem in plain words and gives a next step. Drata is a new connection.
- Optimize shows savings per month again, not a value about 12 times too high.
- Automations, webhooks, and incident investigations wait for your approval instead of declining the action, when their Approval mode asks a person. The request shows in the notification bell. Automations and webhooks default to Auto · Rules + AI · Refuse, so they wait only after you change the Automations row under Run types.
- Triage in a workspace with no credits waits and sends admins one notice a day.
- Only organization Owners and Admins can read audit logs, and share links check permission.
- A turn that pauses for approval is charged once.
- PostgreSQL connections start again.
Connect Claude to your workspaceConnect Claude (claude.ai, the Claude desktop app, or Claude Code) to CloudThinker in one consent step through the Claude Connector. Claude can read your incidents, Optimize, Cyber, Review, and automations, and can start agent runs that follow your Approval mode. Revoke a connection in Account settings → Connected apps.
- Resolve: The triage agent decides which incident an alert joins and whether to page. An unconfirmed signal opens a quiet investigation first, and an alert storm feeds one triage run. Download the root cause analysis as a PDF.
- Review: Thread closes, scorecards, and notes survive a Git provider outage. A Layers tab walks through a merge request in named steps, and a shorter scorecard decides the merge.
- Cyber: Security Testing creates its own test accounts and signs in with TOTP. Findings open on To do, In progress, Fixed, and All tabs, sync to your issue tracker, and export as a report.
- Chat: Steer, which sends your queued message into a running reply, takes effect the moment you press it. Replies arrive one finished block at a time. Search conversations with filters, and message a running subagent from Agents.
- Google Chat: Google Chat adds a guided install, per-space health, link previews, approve and reject buttons, and live progress.
- Automations: A probe trigger runs a cheap check and starts a run only when the check fires. An automation can continue one conversation.
- Optimize: Cost Health leads with what raises your score. Cost Health and Findings download as PDF, and Azure shows the full Cost Health view.
- Connections: Wazuh Cloud and self-hosted share one connection and support several instances. FPT Cloud agents manage volumes, snapshots, and power schedules.
- Pulse keeps unrelated alerts apart, and forwarded log lines no longer open incidents on their own.
- Webhook owners can set their own rate limit. A burst of alerts gets a clear retry answer instead of an error.
- The Auto Approval mode asks for approval less often. Approving a command after an investigation continues the chat.
- Scheduled Keeper runs start again, and older automations can be edited again.
- Elasticsearch connections accept a username and password, and self-signed HTTPS.
Manage your workspace sandboxesOpen a sandbox from the workspace header to see its activity and settings. A crashed or stuck sandbox is replaced automatically, so your workspace keeps working.
- Optimize: Cost Health now includes Reserved Instance and Savings Plan commitments. The weekly digest leads with spend, savings, and the top action. Search and filter Optimize findings.
- Review: Merge an MR or PR from the Review page as yourself. Each MR row shows its diff size.
- Cyber: Reverify a Security Testing finding in one click. The testing agent can log in to your app with saved test credentials.
- Resolve: The Incident Room opens with a short brief. The root cause analysis is shorter and states the cause once.
- Chat: Agents can read your earlier conversations and uploaded files. The PDF preview adds zoom, page jump, and search.
- Notifications: The bell updates in about one second and shows approvals as they arrive. The inbox has filters and day sections. Automation owners hear about failed runs on every channel. See notifications.
- Desktop: The Desktop app updates itself and ships as a Linux
.debpackage. - Connections: FPT Cloud answers questions across your whole tenant without a VPC. GitLab token errors explain the cause.
- Pulse Memory saves cluster memory again, and clusters no longer show as unavailable after a Pulse run.
- Keepers pause in a workspace with no credits and resume when credits return.
- The Azure idle load balancer, idle VPN gateway, empty Application Gateway, and GCP Artifact Registry checks work again.
- When a log doesn’t record the field you ask about, the agent says it can’t tell from that source instead of answering “none”.
- Long background commands run to their deadline instead of stopping after 2 minutes.
Run agents on your own machineRun a chat on a machine you control with a Worker, a self-hosted sandbox. The agent works in the one folder you name and reads its
AGENTS.md from the first turn. Only you can use a personal Worker, while a shared Worker is open to everyone in the workspace. Workers are in beta.- Chat: Switching chats takes under 1 second instead of 2.4 seconds, and streamed replies no longer stutter. Charts zoom, expand, and export.
- BYOK: The organization Owner can pick the model for each part of CloudThinker on your own Amazon Bedrock or Anthropic key. An expired key falls back to the platform model instead of ending the reply. See BYOK.
- Review: Each MR gets one notification row. A Pipelines page lists your pipeline runs. Findings on unchanged GitLab lines now post.
- Resolve: Ask CloudThinker in the Incident Room. Slack and the Room share one root cause analysis. Incident updates on Slack, Microsoft Teams, and Google Chat say what happened, what was found, and what to do next.
- Rooms: Rooms add pins, visible edits, and a jump to any date.
- Optimize: Ask CloudThinker from Cost Explorer, with Ask why on each service. Run a finished full-account scan again from the Overview. Scans now check every billed service by its cost driver.
- Cyber: A Security Testing finding shows the API request and response that prove it, in the app and in the PDF report.
- Guardrails: One page holds ten editable data classes with an organization baseline. A custom pattern takes three fields, and fewer IDs and checksums get masked by mistake.
- Notifications: The bell shows notifications from every workspace and organization.
- Connections: FPT Cloud is a new connection. A failed write says whether a retry is safe.
- A command that asks for several cloud connections runs with the healthy ones and reports the rest, instead of failing.
- Azure DevOps reviews follow the pull request’s latest iteration, and Bitbucket and Azure DevOps review threads resolve correctly.
- Optimize opens an Azure account in seconds instead of up to three minutes.
- A long conversation summarizes its earlier messages instead of growing until it hits the model limit.
Rooms become your team’s shared spaceThe Rooms rail shows your mentions, your Rooms, and the threads you joined. One search field covers every Room, and reactions appear live for everyone.
- Review: Advanced Review keeps its findings when a run crashes. A workspace can reorder, turn off, and write its own review summary sections.
- Chat: A new conversation is now the home page, and a conversation reopens on the model it last used. The
/picker is one ranked list that starts with the commands you run most. - Optimize: Cost Explorer groups spend by service, region, or account. The chart, breakdown, and CSV export follow your choice.
- Cyber: Cyber follows your uploaded PDF or DOCX template when it writes a report.
- Resolve: Connect an incident platform in one step from the Integrations page.
- Connections: Two new connections: Kong Konnect and Railway. An agent can use several AWS and GCP connections in one command.
Related
Quickstart
Set up a workspace, add a connection, and run your first request.
Glossary
Look up the terms these notes use, such as Rule, Pulse, Worker, write, and Cost Health.
Approvals
See how Rules, the Approval mode, and approvers decide each agent write.