Changelog

Skipped triggers now show their reason, and the history is paginated so you can trace back further.

Your agent posts at most one question card per run, and you can keep typing in the composer while it waits for your answer.

When a preview crashes in the builder or an agent chat, it recovers on its own instead of leaving you stuck.

The agent meeting page shows recordings again, the app builder opens from editable app rows, and deployed apps in Slack no longer show duplicate builder links.

Activity now shows note actions and changes, and which skills an agent loaded during a run. Personal skill names stay off the public card, and loading a skill no longer delays the run.

Settings and access links open the right page, integration failures show the service's own error text instead of a generic one, and a no-access reply tags the agent's manager so someone can fix it.

The app builder opens again from app cards, report onboarding finds your data and previews Slack posts correctly, and adding Vybe manager no longer messages anyone.

Agents now know your live Slack timezone, mention you when a reminder or follow-up lands in a shared thread, write their own first message, and keep replies dense. When people are just chatting in an agent thread, the fallback triage stays quiet.

Ask an agent for a password or API key in a message and it will decline, and tell you why. If you paste a secret by accident, it will not use it, repeat it, or store it.

Fixed this week: sign-up links breaking on a stray character, private Slack channels showing up where they should not, the thread not reopening after you connect an integration, and a redundant loader on the way to creating an org. Private agents also now default to executive-assistant settings.

Notes keep their descriptions up to date automatically, which is what makes search results accurate. The description you get when searching is now the one that matches the note today.

When an agent pauses to wait for something, it no longer posts about it in your channel. It speaks when there is something to say.

Organizations are no longer charged for the automatic description checks on notes. One less line on the bill.

If someone mentions a private agent in a channel it is not enabled for, the mention now gets a reply instead of vanishing.

Messages in the app builder that were disappearing now come back as they should.

Before an agent starts a task, it now picks the relevant skill and pulls in only the integration tools that task needs. Less noise, faster responses.

Big file uploads used to fail with a cryptic payload error. Attachments are now stored separately and only the latest message travels with each request, so heavy chats keep working.

When a tool call hiccups but recovers, you no longer get a scary red error icon. Successful Slack tool calls also stop flagging warnings they did not earn.

Pasted text keeps its formatting from one message to the next. What you paste is what you send.

The sign up and login pages got a visual refresh. Same flow, cleaner look.

New chats use Auto by default, with a balanced fallback when a model is unavailable. You can still pick a specific model, but you shouldn't have to.

Half-written messages, attached files, @mentions: each conversation now remembers its own. Switch between chats without losing what you were typing.

When you connect an agent to Slack, channels are now visible, searchable, and preselected where it makes sense. Less scrolling through an unsorted list.

The agent sidebar got a redo, and conversations now title themselves from your first message until you name them. Small thing, big difference when you're hunting for last Tuesday's chat.

Exported PDFs keep your artifact's design, the app editor keeps its full height, settings no longer freeze when you navigate away, and GPT-5.x strict-mode tool schemas translate correctly. Live queries also accept database names.

Images your agents generate post as standalone Slack messages instead of getting lost in threads. Tool calls also run through Slack's interaction system, which makes progress updates steadier.

Routing now prefers GLM 5.3 Flash providers that serve cached tokens. Same model, less waiting.

Two fixes to interactive questions: they stay put when you switch chats and come back, and agents can rely on them again.

The redesigned layout landed this week. Scroll no longer jumps mid-conversation, which on its own might be the best fix here.

Chatting with an agent while it replies no longer causes the page to jump around.

Chat errors used to flash and disappear before you could read them. They now stay visible until you dismiss them, with calmer styling for expected issues like hitting a usage limit.

Fixed a handful of bugs where an app preview would get stuck loading after a brief outage instead of recovering on its own.

GPT-6 Astra, Gemini 3.8, Qwen3.8 (Max and 27B), and Muse Spark 1.3 are now available in the model picker.

Chat no longer jumps around mid-response. Text stays put while the agent is still typing.

When a run gets delayed, you can now see it queued up and choose to let it ride or run it now.

Notes with an over-long description used to fail outright. Now they save, just trimmed to fit.

When a tool call fails, you will see the actual error now instead of a placeholder that just says an error occurred.

A cut-off emoji in a message could freeze an agent channel entirely. That is fixed now.

Redshift connection configs were showing sensitive fields that should have stayed masked. That is fixed now.

Calls to app APIs with query strings were sometimes dropping the parameters and returning a garbled HTML error instead of the real one. Query strings now come through correctly.

If an MCP integration's token stops working, it is now marked as expired instead of failing silently, so you know it needs reconnecting.

If a preview hits a chunk load error while you are building, it now reloads automatically instead of leaving you stuck on a broken screen.

The sidebar user menu now shows which organization you are in. Useful if you manage more than one.

Dropped the Slack profile picture step from onboarding. It was not adding value, so it is gone.

Opening the model selector no longer crashes the chat interface.

Added Slack permissions that were missing, which were causing errors for some integrations.

Fixed a bug where switching models in chat could crash the whole interface.

Fixed two separate bugs that made GPT-based agents fail after using a tool or after being switched to a different model mid-thread.

Fixed an issue where email attachments were being sent to the agent as raw text instead of readable files.

Removed a crypto-related warning that was popping up for no good reason.

Expanding a failed tool call shows the actual error instead of a generic failure message.

The capabilities panel got a visual overhaul, making it easier to see what an agent can do and where it still falls short.

The sidebar is now resizable, shows status spinners while agents work, and lets you queue messages across threads without losing your place.

When multiple agents share a Slack channel, the right one now responds. Previously the wrong agent could jump in.

Switching organizations now lists every org you belong to, instead of a partial list.

Claude Opus 5 is now the default smart model, and GPT-5.6 Luna, Gemini 3.6 Flash, and Kimi K3 are all available to use.

We tuned how generated apps handle initial rendering. New apps feel snappier the moment they open.

Apps without a production deployment can now be archived directly instead of sitting around in your workspace.

Fixed duplicate Slack cards, a mislabeled empty-channels state, and a sync issue that could hide agent tools mid-conversation.

Redesigned the publish popover and button states. Clearer at a glance what stage your app is in before it goes live.

Create apps and add capabilities without leaving Slack. One less tab to juggle.

Slack now supports interactive questions. An agent can ask you something and get your reply without switching to the web app.

It is collapsed by default now, you can toggle it open or shut anytime, and it remembers whichever state you left it in.

Agent chats and Slack conversations get a full 30 minutes before timing out, so longer tasks do not get cut off mid-run.

Magic links and OTP codes are good for 10 minutes now instead of 5.

Agent run history now shows the real tool names instead of generic placeholders.

Fixed a crash when an agent searched meetings, and a Slack bug that broke mention-only routing for routed agents.

When you give your agent access to an integration account, it turns on automatically. No extra step to flip it on yourself.

Close the free-trial start banner once and it stays closed for your whole org.

Some agents were silently ignoring Slack events. That's fixed, so your agent will see and respond to messages again.

Fixed cases where trigger setup showed confusing or broken states for options that depend on other fields, plus a bug affecting agent tools tied to triggers.

New agents start on Sonnet 5 unless you pick something else.

When agent permissions are off by default, you can now see exactly which databases and integrations are restricted in the agent's permissions summary. This makes it easy to see which tools need an explicit grant.

Agent triggers now require a connected integration account and are automatically cleaned up when that integration is removed. This prevents dangling or duplicate trigger executions.

Fixed an issue where interactive connection cards, like the Gmail connector, would render collapsed inside the agent's internal thinking blocks. They now show up as clear action cards in the main chat body. Note and skill wikilinks in chat are also clickable deep-links now.

Fixed an issue where saving notes would automatically switch your active panel back to the Memory tab, keeping your focus on your active workspace instead.

We stopped loading heavy bootstrap prompts on every single conversation after the first thread. Pruning old YAML setup questions and refining progressive disclosure prompts makes follow-up chats feel much faster.

We fixed a database connection bug where undefined SSL options shadowed the default sslmode. We also resolved TLS server-auth fallbacks, fresh agent creation errors, and added step limits to cron agents to stop runaway runs.

Pass-or-fail grading was too rigid to show real progress. Evals now score test runs by averaging criteria, alongside new per-agent default model choices for evaluation runs.

The in-progress reasoning view got a redesign: better sizing, a clearer handoff to the actual response, and a usage indicator so you know what the agent is working on.

Granting an agent access to an integration now auto-proceeds instead of waiting for a manual click. The grant-access card is also cleaner.

Credit figures throughout the UI now show the USD equivalent inline. Removes the mental math.

Cron agents were blocked from calling app API tools due to a missing permission check. Fixed.

/api/calls was returning full multi-megabyte transcript payloads on every request. Now it returns a trimmed response; full transcripts fetch separately when needed.

Dropped connections no longer mean lost responses. The platform now resumes or reconciles persisted messages when a stream cuts out, so an interrupted run picks back up cleanly.

The askUserQuestion tool is no longer exposed to agents running in Slack channels, where it was creating unnecessary prompts.

Loading states, thread messages, markdown rendering, onboarding menus, usage dashboard, and card layouts all got fixes this week.

SQL errors from runSql and agent databases now show the actual Postgres or Neon error, not a generic fallback.

The Slack panel setup content is cleaner. First-time connection feels like onboarding instead of reading documentation.

Connecting a custom sending domain has fewer requirements now. Less friction when setting up agent email.

The agent integration settings dialog is reorganized and easier to scan when managing multiple integrations.

Two issues patched: agents now favor DMs correctly when a channel message is denied, and Slack actions only mark as sent after the tool call actually succeeds.

Credit allocation and trial day defaults were wrong on certain paid plans. Both are fixed.

Creating notes from inside an agent session was failing in some cases. Works cleanly now.

Two agents can no longer share the same inbox address. The system enforces uniqueness at setup, so you catch the conflict before it causes problems.

Tool categories in the capabilities panel are labeled now. Easier to find what you are looking for without having to open each section.

After connecting Slack, agents now open the relevant subthread automatically. There is also a new tool that lets agents surface the Slack install button directly in conversation.

Creating a cron job that already exists used to fail silently. Now you get an explicit message so you know what happened.

Fixed a visual glitch where the note link popover would flash when switching tabs. Small fix, but noticeable.

Settings sidebar now has sub-items so you spend less time hunting for the right page.

More reliable enforcement of which senders can trigger your agent by email.

Cleaner layout, tighter grid, and the creator and last-edited metadata is gone.

New agents get an avatar that matches the gender of their name. Small thing, but it matters.

Active meetings now show a pulse in the agent interface so you know a call is being recorded.

Agents can now BCC recipients when sending emails.

Tighter controls over which contacts can trigger an agent via WhatsApp.

Each scheduled task can now run on a different model. Useful for cost and quality tradeoffs across jobs.

When web search providers fail, agents now receive structured fallback guidance instead of a dead end.

Slack responder is now blocked when agent access is disabled. Mention-only response mode is available.

Integration and custom API secrets are now invalidated immediately on soft-delete.

Cron descriptions can reference skill slugs — skill content expands at runtime so updates flow through automatically.

Slack read tools now return the channel name or DM partner handle alongside messages.

Pressing Escape in nested modals now closes only the innermost one.

Fixed a bug where prices checkout subscriptions were incorrectly auto-canceled.

MCP servers are now subject to the same access control and permission checks as other integrations.

App sandboxes inactive for 29 days are automatically expired to free up resources.

Multiple fixes to sandbox process timeouts, error handling, and H2 client timeouts for more reliable code execution.

The scheduled task limit jumped from 10 to 50 per agent. If you were hitting the ceiling, you are not anymore.

You can now stop an agent mid-stream. No more waiting for a runaway response to finish.

WhatsApp is now available as an agent channel via Kapso API. Agents can send and receive WhatsApp messages.

Action inputs are now validated against configurableProps at runtime before any external call is made. Catches bad inputs early.

You can now favorite agents. Favorited agents show as cards, same visual treatment as apps.

Email, calendar, meetings, Slack, and phone are now consolidated into a single agent communications page.

Agent notes now live on their own page with direct URL navigation. Easier to find, edit, and share.

Scheduled task descriptions render markdown, including inline note links.

Chat, heartbeat, and SMS all run through the same core execution engine now. Behavior is more consistent and easier to test.

You can now give integration accounts human-readable names instead of relying on default labels.

Granola is now available as a native integration. Meeting notes flow directly into agent context.

We cut the Nango dependency entirely. Integrations now run on our own stack, which means fewer moving parts and no more Nango-related auth issues.

Context compaction is live, along with proper reset behavior on new top-level messages. Heartbeat step limit also raised from 10 to 30.

The preview navbar now has a fullscreen toggle. Small UX thing, but agents that build apps needed it.

Agents in Slack now stay assertive. The previous behavior where they could loop into clarification questions is gone.

We've modernized the interface with a streamlined sidebar, improved transitions, and updated icons for a cleaner, more intuitive experience.

You can now configure trusted SSO domains, making it easier and more secure for users to sign in through approved identity providers.

SSO enforcement has been added, requiring all users to sign in using Single Sign-On.

Made additional enhancements to the workflow interface for smoother navigation and usability.

When adding a custom API, Vybe now automatically validates your OpenAPI specification. This helps catch errors early and ensures your API integrations work reliably.

Resolved a problem that prevented invitations from being created correctly. Users can now send and receive invitations without errors.

Enhanced the updatefile tool to provide clearer and more actionable feedback during file updates.

The Integrations page now features a new Table View for each section, making it easier to view and manage your integrations.

You can now softly delete apps instead of permanently removing them. This feature helps prevent accidental loss by allowing recovery of deleted apps.

Users now receive clear error messages if sending an invitation fails.

The editor and chat interface have been updated with improved background colors, cleaner gradients, and subtle borders for a more polished look.

We've moved our sandboxes and Neon projects to be geographically closer to the core vybe application. Expect faster performance on deployed apps.

The sandbox preview iframe now allows clipboard writing, resolving errors when user apps try to use the Clipboard API.

Enhanced the look and feel of the Editor Tab Bar by refining its hover and selected states for better visual feedback.

Messages from the AI now fully support Markdown formatting.

Resolved a bug that prevented viewers from accessing deployed applications.

Updated the look and feel of feature suggestions.

The Data panel now features a new Integrations tab along with simplified horizontal tab navigation. This makes it easier to view and manage all your connected services and databases in one place.

Chats now load more quickly by displaying recent messages first and loading older messages as you scroll. This improves performance, especially in conversations with many messages, so you can start chatting sooner.

The chat tool now includes an AI-assisted type checking feature, offering quick type verification without needing a full build. While it's currently slower for larger apps, this update helps catch type errors earlier to streamline development.

Tool outputs that return JSON are now shown in a clearer, easier-to-read format thanks to a new reusable UI component. This update also fixes imports in the Gmail Inbox component for better reliability.

We've improved the quality of AI-generated suggestions to provide more accurate and helpful guidance. This update helps you work more efficiently and make better decisions within the app.