August 2026

Build: 0.11.4 Commit: 935834be9b Built at: 2026-08-18 22:42:29 UTC

The one where your assistant can finally see what you're pointing at, and the sidebar got a haircut.

• Show your assistant what you're looking at. During a voice call, you can now open the camera and share what's in front of you. Point at a broken part, hold up a document, or show the screen you're stuck on. The assistant sees the frame and talks through it with you in real time. No more describing what something looks like when you can just show it.
• Speak in any language, and pick which one the assistant listens for. Speech recognition now defaults to multilingual, so you can switch between languages mid-conversation without the assistant losing the thread. If you want to lock it to one language for accuracy, there's a new listening-language picker in Settings → Voice. Talk Italian, switch to English for a technical term, and the assistant follows along.
• Each onboarding avatar now has its own voice. When you're picking your assistant for the first time, every avatar speaks with a distinct voice so you can hear the difference before you choose.
• Subagents are now researcher, builder, or advisor. The old subagent types are consolidated into three clear roles with platform-backed guarantees about what each one can do. Profile isolation and loop protection are built in by default, so a subagent stays in its lane and does not run forever. You get the same three choices whether you spawn one manually or a schedule fires one.
• UI update on personality. The radar chart that plotted personality traits is gone. In its place is a cleaner signature mark that captures the same character without looking like a flight instrument.
• Guided setup for Telegram and Slack. Both channels now have a step-by-step walkthrough that handles the connection flow for you. The Slack wizard is back over Slack's own setup page, and Telegram has a new walkthrough that does the work the old setup skill used to do manually.
• Plugins can update themselves. Plugins now support opt-in hourly auto-updates. When a plugin author ships a fix, you get it without reopening the marketplace. Turn it on per plugin, and the monitoring worker handles the rest.
• Pin a schedule to a specific inference profile. Every schedule can now lock to a concrete model and provider, so a 6 AM summary always runs on the profile you chose for it. Before deleting a profile, you can see and move the schedules that depend on it, so nothing silently breaks.
• Skill revision history. The skill detail page now shows recent revisions, so you can see what changed and when. If a skill update shifts behavior, the history tab is where the diff lives.
• Sidebar UI polish. Labels are renamed to be shorter and clearer. The grouping tabs are replaced by a single dropdown, so you switch sections without hunting for a tab. The collapsed rail is now a clean column of circles, and section headers have their own cards instead of blending together.
• The billing page got a redesign. Your current plan and what's next sit side by side as tiles instead of a long table. A custom plan configurator is available if you want to tune cadence, limits, and baseline machine without emailing anyone.
• Daily credit limit with a proactive banner. You can set a daily credit limit, and if auto top-ups are on, a limit is required. When you're approaching the limit, a banner tells you mid-turn instead of waiting for the next failure. A deep-link anchor jumps straight to the setting.

The one where voice mode gets more useful across devices, files stop hiding in chat, replies can recover from interruptions, and the sidebar finally feels calm.

• Voice gets better across devices. Voice conversations now support richer controls and clearer status across web and iOS. You can minimize or reopen the voice room, mute or end a call, approve or deny actions from the iOS Live Activity, and follow an active voice session from iOS notifications, launcher shortcuts, or Quick Settings.
• Preview your files in chat: Images, PDFs, spreadsheets, and text files can now appear as previews or file cards right in chat.
• A faster model choice: Speed is now its own managed profile for quick replies, while Balanced uses GPT-5.6 Luna where available.
• Credit warnings tell you what to do next: If you are running low or out of credits, Vellum explains the problem in chat and gives you a clear way to add more.
• App search gets to the useful part faster: Matching text now appears closer to the front of each result, so you can scan long conversations without digging.
• Sidebar reorganized: Start with one simple view, switch to grouped conversations when you want more structure, resize pinned content, and see clearer labels while everything loads.
• Interrupted replies can continue: If a response stops halfway through, your assistant can pick up the turn instead of making you start over.
• Variety of bug fixes: Archive All asks before clearing conversations, compaction repairs broken tool results, macOS downloads save correctly, push notifications can reach you after you leave a chat, and Slack mentions keep their formatting.

This update makes Vellum easier to use every day: talk to your assistant from your iPhone, keep your conversations organized, connect more services, and let your assistant remember more without exposing sensitive information.

• Your assistant is easier to reach on iPhone. Talk to it from the lock screen, while using another app, or with Siri and the Action Button. Live Activities show what is happening without making you reopen the app.
• Keep your conversations organized. Create groups for work, projects, trips, or anything else. Drag chats into a group, give it an icon, and arrange your conversation list the way you want.
• Connect plugins more safely. Plugins can now receive information from other services after you approve the connection in settings. You stay in control of what gets connected.
• Your private information gets a warning. If you accidentally paste an API key, password, or other sensitive detail into a message, Vellum warns you before sending it to an AI model.
• Model settings is much easier to understand. Model choices, provider connections, and assistant preferences now open in clear side panels instead of being buried in layers of menus.
• More ways to search and talk. Keenable joins Brave and Tavily for web search, while GPT-5.6 Terra and Luna give subscribers more model choices. Luna is tuned for faster voice responses.
• A cleaner, smoother app. The sidebar, iPhone chat screen, tips, sign-in flow, Telegram setup, Slack links, conversation recovery, and scheduled tasks all received fixes and polish.
• More possibilities for plugins. Plugin developers can now build tools that do more, search their own information, and store structured data, which opens the door to richer assistant experiences.

The one where voice mode keeps up, Discord becomes part of your assistant, image generation finds a proper home, and Vellum gets better at protecting the things you paste into chat.

• Voice mode keeps up with the work: Hear updates when something actually changes instead of getting the same message every few seconds. Minimize the voice room without leaving the call, see clearly when your assistant is listening or speaking, and stop it when you need to.
• Image generation gets a proper home: Choose image generation as a provider, just like you choose a model. It is no longer hidden inside a separate mode.
• More models to choose from: Claude Opus 5, GPT-5.6 Sol, Kimi K3 through Fireworks, and Poolside are now available for different kinds of work.
• Get help without leaving chat: Type /doctor to open Vellum’s diagnostic tool with your first message already filled in.
• Files are easier to open: File paths in chat become clickable links, and Vellum asks for permission when a task needs a file outside your workspace.
• Everyday polish: Billing works better on mobile, background work can finish before you shut down with vellum sleep --wait, rotated API keys recover automatically, and notifications stay connected to the conversation they came from.

The one where Voice Mode goes live, Vellum becomes its own web search provider, the memory graph and bookmarks graduate to everyone, and the billing page gets a full restructure.

Voice Mode, live. The most delightful, real agentic AI voice mode you'll ever use! You can now have real-time voice conversations with your assistant. Tap the mic button in the composer, start talking, and your assistant responds out loud in real time. Read more about voice mode from the creator itself: https://www.vellum.ai/blog/vellum-assistant-ai-speech-pipeline-explained

Vellum is now a web search provider. Web search used to be a "mode" your assistant toggled on and off. It's now a proper provider, sitting alongside Google and Tavily in the search provider catalog. Pick it once in settings and your assistant uses Vellum's managed search for every web query. Simpler config, cleaner architecture, same results.

Memory graph. The visual map of how your assistant's memories connect to each other is now available. Open the Memory tab and click the graph icon to see a live, interactive network of everything your assistant knows and how it all links together.

Bookmarks. Bookmark any message in any conversation and find them later in the sidebar.

Billing, simplified. We reworked the billing plans to be less confusing. Instead of the old Base and Pro tiers with a dozen machine and storage combinations, you now pick from four clear plans: Free, Mighty ($30/mo, recommended), Super ($100/mo), and Ultra ($200/mo).

Each plan bundles a computer size, storage, and credits into one price, so you know exactly what you're paying for. Need something specific? The Custom Plan option lets you tailor CPU, RAM, storage, or tokens to your needs.

New Gemini models. Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are now available through the Google provider. Select them in your model profile like any other model.

Archived conversations show up in search. Search results now include archived conversations, not just active ones. Find that old conversation you tucked away without un-archiving it first.

Assistant avatar feels more alive. The assistant avatar now strains while it's working and grows when it lands a response. The assistant overview page got a blurred photo backdrop with glass cards. The provisioning takeover (the animation that plays when you upgrade plans) is tinted with your assistant's own avatar colors, and it no longer flashes a stranger's avatar while yours loads.

Open in Slack. The conversation header menu now has an "Open in Slack" option that jumps straight to the Slack thread where that conversation is happening. Useful when you need to check the raw Slack context for a conversation that's been running in the app.

The one where Skills and Plugins become one page, your assistant gets a voice you can actually pick, and the billing page grows up.

My Superpowers. Skills and Plugins used to live on separate pages with separate vibes. They're now one unified page called "My Superpowers" so you can browse, install, and manage everything your assistant can do without bouncing between tabs.

Pick your assistant's voice. Managed TTS now has a voice picker. Browse available voices, preview each one before choosing, and see which provider it comes from. Your assistant's voice is no longer whatever the default happened to be.

Kimi K3 on OpenRouter. Kimi's K3 model is now in the OpenRouter catalog, so you can select it as your assistant's model without any extra setup.

Retry button on the latest turn. If your assistant's last reply wasn't great, there's now a retry button right on it. One click, new response. No need to retype your message or start a new chat.

Email attachments land in your workspace. When someone emails your assistant and the message has attachments, those files are now ingested into your workspace automatically. Photos, documents, whatever they sent, your assistant can see it and work with it.

Plugins install their own dependencies. When you install or upgrade a plugin, it now pulls in whatever dependencies it declared. No more manual npm installs or broken plugins because a package was missing.

One profile for all actions in Action Overrides. When configuring Action Overrides, you can now apply a single model profile to every action at once instead of setting each one individually.

Library moved to the Assistant page. The Library tab relocated to the Assistant page, and the navigation's assistant cluster got rebuilt to be cleaner. Fewer top-level tabs, more logical grouping.

Schedule timezone defaults to yours. When you create a recurring schedule, it now defaults to your timezone instead of UTC. No more waking up to 3 AM nudges because the scheduler thought you lived in London.

Download button on media previews. Images and videos in your workspace now have a download button right in the preview header. One click, file on your machine.

The one where Settings gets reorganized, plugins can ship their own apps and your assistant shows off the skills it taught itself.

• Plugins can ship apps. Plugins can now bundle full apps that open right in your library. You can upgrade plugins installed from a GitHub URL directly, and external installs now land with a confirmation deep-link.
• Your assistant shows off the skills it taught itself. When your assistant writes a new skill through its own memory (the retrospective process that turns patterns into reusable skills), it now shows a clean card in chat with a one-sentence summary of what it learned. No more generic "skill created" header, just a readable line about what the skill actually does.
• Settings, reorganized. The whole Settings page got rebuilt. General is now split into clean Profile, Version, and Appearance sections with 2FA and a Danger Zone. The old Security and Keyboard Shortcuts tabs are gone (shortcuts moved to a Preferences modal). There's a new Debug tab where you can go to restart your assistant or access its terminal. And if you're on a cloud assistant, the Credit-usage chart got moved to the Billing tab alongside your token breakdown.
• Providers, simplified. Pick a provider and Vellum figures out the auth. The Providers modal is now provider-first (no more choosing a "connection type" or "auth type"), and your profile picker shows a single Vellum entry with the upstream model derived automatically. The legacy LLM resolution layer is gone.
• Baseten is available as a BYOK provider. (bring your own API key) serving Thinking Machines Inkling, with reasoning-effort passthrough.
• Assistant overview, rebuilt. The assistant page is now a fun and interactive screen that represents your assistant personality.
• Email attachments. You can now send file attachments in outbound emails on both Outlook and the native emailer.
• "Summarize up to here" in chat. The per-message "Summarize up to here" action is no longer behind a feature flag. You'll see it on every message, with a progress indicator while it works.

The one where Vellum starts handling your speech for you, GPT-5.6 lands with prompt caching, and your phone learns to swipe.

• Vellum can handle your speech now. Speech-to-text and text-to-speech used to mean going and getting your own API keys from a provider like ElevenLabs or Deepgram. Now Vellum offers speech through its own managed service, and it turns on automatically when you connect your Vellum account and deducts from your credit balance. You can still use your own API keys any time from the Text-to-Speech and Speech-to-Text cards in Settings.
• GPT-5.6 with prompt caching. GPT-5.6 models are now available. They support prompt caching, which means the big chunks of text that get sent every turn (system prompts, tool definitions, your conversation history) get cached by the provider, so follow-up turns cost less. You also see a visible "thinking" summary when the model reasons before answering, instead of a silent gap where you're wondering if it's working.
• Touch gestures on mobile. Swipe left on a conversation to reveal delete, pin, or rename. Press and hold a message to get copy, edit, or retry. Swipe through a message's image gallery instead of tapping little arrows. Swipe app windows and drawers away to close them. Plus a stack of iOS fixes: the keyboard dismisses properly when you open the mobile menu, swipe-action rows don't accidentally open the menu anymore, and the light-mode seam at the bottom of the screen is gone.
• Plugin marketplace is more reliable. Before, you'd sometimes see a plugin in search results that 404'd when you tried to install it. Search, install, and details now all pull from the same source, so that doesn't happen. Plugins can now include their own mini web pages (like a dashboard or a settings panel), ship proper image icons instead of just emoji, and the CLI fills in a missing package.json for you when you're scaffolding a new one.
• Channel trust floors. There's a settings card on the Privacy page that lets you control who can send messages to your assistant on each channel (Slack, Telegram, etc.).
• More capabilities for apps. Apps running inside Vellum can now push messages into conversations, include bundled files like images and fonts, and receive live updates when something changes externally.
• Smoother streaming chat. Hover actions on messages now collapse into a clean avatar reveal instead of cluttering the screen. Streamed text fades in instead of popping in, and the first-turn jump where the UI would briefly glitch is fixed. The credential card in Settings is redesigned with a click-to-reveal toggle so you don't accidentally expose your API keys. Links inside rendered components (like recipe cards or search results) are now clickable. The risk section in the tool drawer is renamed to "Risk Level" and the old trust-rule button is gone.
• New skill: chat-complex-documents. A bundled skill for turning messy documents (PDFs, scanned files, Word, Excel, PowerPoint) into clean structured text you can search and ask questions about. Available in the skills catalog.

The one where you pick a default AI provider once and forget about it, your images and attachments work properly, and your assistant's mobile app gets a full design polish.

Pick a default provider and stop thinking about it. Mark one provider connection as "default" and your assistant always tries that one first. If it's unavailable, you see a clear reason why instead of a cryptic error. This replaces the old system where the assistant guessed based on rules most people never set up.

Errors that make sense. When a model call fails, you get a plain reason like "rate limited" or "invalid API key" instead of a status code you have to Google.

New models and providers. GPT-5.6 is available through OpenAI, and Grok 4.5 is available through OpenRouter. You can also route through the Vercel AI Gateway as a new provider called Atlas Cloud.

A better mobile experience. Edge swipe opens the menu and goes back on detail pages. iOS gets a native reply option in the text-selection menu. The side menu scrolls smoothly with floating New Chat and Preferences buttons. A full design review polished spacing and alignment throughout.

Images and attachments work properly now. External images render inline. Images your assistant generates are openable and downloadable. Videos play in-app. Everything survives page reloads instead of disappearing.

See what your scheduled tasks actually did. Open a schedule run and you can see the real conversations it produced, not just a status. Failed runs are now marked as failed instead of showing as successful with empty output.

Better CLI. Manage model profiles and providers from the terminal, see per-schedule cost breakdowns, and install marketplace plugins faster. CLI memory usage dropped by roughly two-thirds.

Your assistant recovers from crashes. If the process dies mid-conversation, it picks up where it left off on restart. No more stuck "thinking" states.

13 new plugins in the marketplace. Google Drive Search, Cofounder, Coffee Aficionado, Plant Doctor, Tennis Companion, Kitchen Companion, SMB Inbox Brief, Family Briefing, Personal Finance, Proactive Slacks, Cognee Memory, Influencer Marketing, and Vellum Client QA. Plus a Gmail Trigger skill (beta).

The one where onboarding gets a personality, your calendar skills learn to juggle multiple accounts, and the Channels page gets a much-needed makeover.

• A better onboarding. New users now answer a few questions about their work, preferences, and the personality they want before meeting their assistant. So the first conversation starts with an assistant that already knows something about you, instead of a blank slate asking "how can I help?"
• Work and personal calendars, together. Connect work and personal accounts (Google or Outlook). Your assistant now knows which is which and picks the right one for the task, instead of guessing and sometimes getting it wrong.
• Channel permissions got stricter. Permissions for channels (Slack, Telegram, etc.) are now checked every single time your assistant tries to use a tool, not just once when the channel connects. If a channel isn't allowed to do something, it gets blocked at that exact moment.
• A cleaner Channels page. Each channel now has its own tab in settings instead of everything stacked in one long list. Find what you need without scrolling past what you don't.
• Pick which Slack channels your assistant lives in. The Slack channel list now shows which channels are active and lets you limit your assistant to only the channels you're a member of, instead of every channel in the workspace.
• New in the plugin marketplace: Motion Design. Adds smooth animations to your assistant's web UI. Install it with assistant plugins install motion-design or by asking your assistant to do install it.
• Faster startup with plugins. Plugin hooks now load only when needed instead of all at once at startup. If you have plugins installed, your assistant boots faster.
• Bug fixes. Piping input to the CLI works reliably across shells. Full-screen views on mobile follow the iOS keyboard properly. Assorted UI component fixes.

The one where setup gets simpler, phone calls get more reliable, and your assistant gets better at remembering what matters.

• One connection for all Vellum-hosted models. Before, if you used the models Vellum hosts for you, each one needed its own connection in settings. Now there's just one. Connect once and every hosted model works. The built-in profiles (Quality, Balanced, Cost-Optimized) also can't be deleted or renamed by accident anymore. As a result, we also fixed a bug that now lets you use platform managed profiles in local assistants.
• Check your credit balance from the terminal. Your assistant can run assistant platform credits to see how much you have left.
• Control who can talk to your assistant. Each channel (Slack, Telegram, Phone) now has its own settings for who can reach your assistant and what they're allowed to do. Look for the new "Who Can Reach" page in settings, and a Channels tab on the About Assistant page that shows everything in one place.
• Phone calls are more reliable. Two annoying bugs are gone: when you interrupt your assistant mid-sentence, it actually stops talking. And calls end properly when you hang up, instead of hanging around. Calls should also feel faster because we rebuilt the audio pipeline that handles them.
• Better memory. Your assistant is now better at bringing up the right context at the right time, without you having to repeat yourself. This is on for everyone automatically. You can also edit or delete individual memories from the terminal with assistant memory items.
• Fewer failed actions. Sometimes the assistant tries to use a tool with a wrong name and the action just fails. Now it gets pointed to the right tool automatically, or gets a "did you mean?" suggestion. Things that used to dead-end now just work.
• Fixes itself when a connection drops. If the background process that handles logins and permissions loses its connection, it now reconnects on its own. Before, it would stay broken until you restarted.
• Slack cleanup. Approval cards no longer repeat themselves or leak the approval code. Replies stream in with proper spacing instead of words running together. And other bots are no longer mistaken for humans in your contacts.
• For plugin builders: hooks got a big upgrade. Hooks can now show things in the UI, not just run in the background. Edit a hook file and it reloads instantly, no restart. And there's a new hook that fires when a conversation is deleted, so plugins can clean up after themselves.

The one where your assistant actually feels like someone, onboarding stops being intimidating, and Slack messages arrive as they're being written.

• New onboarding. No more blank slate. First-time setup is now a guided walk-through: name your assistant, pick a personality, connect your tools, and you're ready before your first conversation. We also do a quick research pass on you, so your assistant gets to know you immediately, your work, your preferences, your context, instead of starting from zero. You go from "what is this" to "let's go".
• More distinct personality. Your assistant's personality now comes through clearly and consistently. Same assistant, same vibe, every conversation. It actually sounds like itself instead of defaulting to generic-bot-speak. The sarcastic one stays sarcastic. The chill one stays chill.
• More succinct responses. Responses are shorter and get to the point faster. If you ask "what's 2+2," you get "4," not a paragraph about the history of arithmetic. Less scrolling, more doing.
• Notification controls on the homepage. The homepage now has "Clear all" and "Mark all as read" buttons. Got 47 notifications piled up from a busy afternoon? One click, gone. No more dismissing them one by one like it's 2015.
• Redesigned Slack setup. Connecting your assistant to Slack has a cleaner, clearer setup flow. Fewer steps, less guessing, no mystery "something went wrong" moments.
• Slack messages stream live. When your assistant sends a message in Slack, you watch it appear word by word as it's being written. No more staring at a "typing..." indicator wondering if anything is happening. You see the thought forming in real time.
• Ollama works for local assistants. Ollama, which lets you run AI models locally on your own machine, now works properly for self-hosted assistants. Want to keep everything on your laptop, no cloud, no API keys? That actually works now.
• Quality profile runs on Fable. The managed Quality model profile now uses Fable as its underlying model. Better outputs on the highest-quality setting.
• Background task interfaces. Subagents, ACP runs, and other background tasks now have dedicated views so you can follow what's happening while they run. ACP (Agent Client Protocol) lets your assistant hand off coding tasks to external coding agents like Claude Code or Codex, which run autonomously and stream results back. So instead of "did it finish? is it stuck? should I refresh?" you watch the work happen in real time.
• Read only conversations are now resumable in the assistant. Previously, conversations started in another channel showed up as read-only, meaning you could see them but couldn't reply. Now you can pick up right where you left off and keep the conversation going without switching back to wherever it started.

A feature release headlined by a memory system overhaul and rebuilt Slack rendering, plus plugins are now generally available.

• Self-improving skills. The assistant now writes its own skills based on memory, edits them, and knows what to do better next time. The retrospective writes and edits its own procedural notes, and recall is sharper for named people and things. The memory worker can be started, stopped, and checked from config.
• Richer Slack UI. Assistants now use native Slack images and tables instead of just sending plain text. Messages render through a real markdown parser, so images, tables, column alignment, and headings all work, and reactions show as emoji.
• Heartbeat on by default. A heartbeat is a moment when your assistant quietly checks in on its own, without you asking, to catch reminders, follow-ups, and anything time-sensitive. It's now enabled out of the box, with up to 10 runs a day. The Schedules tab gets shareable links to individual schedules.
• Improvements to default models. The managed Balanced profile moves to GLM 5.2 with medium reasoning effort. The Frontier profile is retired in favor of Quality. Together AI profiles now show topP and effort controls.
• Chat improvements. Fixed idle scroll flicker, and thinking and tool drawers now stay open after a turn finishes. Thinking previews now stream live as they generate.
• New plugins. You can install plugins directly from a GitHub URL, and the Plugins tab now matches the Skills tab. New in the plugin marketplace: Fitness Companion, Reading Pal, and Writing Coach.

A feature release headlined by the new onboarding flow in the web app, plus MCP server management, plugin publishing, and redesigned subagent and workflow UIs.

• Guided onboarding in the web app. New users get a guided flow after signup that suggests plugins based on how they plan to use Vellum, connects Google Calendar, and schedules a Day 2 check-in.
• MCP server management. MCP settings have moved under Integrations. You can now view connected MCP servers (including which tools are registered on each and the token contribution) and remove servers from the UI. Adding OAuth-based servers (with automatic auth flow) and revoking connected OAuth accounts from the web UI are still in development.
• Plugin publishing. A new assistant plugins publish CLI command lets authors publish to the marketplace directly. The plugins list command gains an -all flag, and installed plugins now activate without restarting the daemon.
• Advisor streams live. The advisor now streams its guidance as it generates, with search and reasoning activity visible in real time alongside the final advice. It receives full conversation context and runs as a blocking step.
• Subagent UI redesign. The subagent detail panel is redesigned with a phase-grouped timeline and a new Active Subagents floating overlay showing running subagents at a glance.
• Computer-use and macOS polish. Selection and cursor key actions are no longer incorrectly flagged as having no visible effect. The macOS app nudge banner now appears 24 hours after first seen instead of immediately. The mac-helper executable gets proper per-environment naming so it displays correctly in Privacy & Security settings.
• Model profile updates. The managed Speed profile moves to DeepSeek V4 Flash via Fireworks. The managed Balanced profile's reasoning effort is lowered to medium. topP and effort controls are now visible for Together AI profiles.
• Quote reply is generally available. The quote-reply flow, previously behind a flag, is now on for everyone, you can reply inline to a specific message and the original is quoted as context, no opt-in needed.
• Model profile picker on schedules. Schedules now expose the model-profile selector, so a scheduled run can be pinned to a specific profile (cost-optimized, balanced, etc.) instead of inheriting the active one at fire time.
• Channels and gateway. Channel trust floors move inline into Contacts → Channels and are enabled by default. Consent version comparison is now monotonic, stopping cross-client re-consent loops. Approval requests now notify the requester and release interaction on expiry.
• Workflows UI. The workflow detail panel is redesigned with an active workflows floating overlay pill. Directory-style saved workflows now resolve correctly.

A maintenance release that fixes a startup migration issue.

• Database migration no longer gets stuck on startup. Two startup migrations could crash and leave the daemon stuck replaying a failing migration on every launch if the database had already been partially updated by a later migration. Both are now idempotent and safe to re-run regardless of schema state.

A feature release that refreshes the model profile lineup, adds image support for text-only models and many stability improvements and bug fixes.

• Quality profile now uses GLM-5.2, plus a new Frontier profile. The Quality profile now points to GLM-5.2. A new Frontier profile has been added for users who want Claude Opus, available as the highest-tier model option.
• Balanced profile now served via Together AI. The Balanced model profile, which runs Minimax M3, has moved from Fireworks to Together AI as its hosting provider. The model itself is unchanged, only the infrastructure behind it.
• Better image handling for text-only models. When a tool returns an image, like a screenshot, the assistant now describes it in words so models that can't see images still understand what's there.
• Fixed errors with newer models. Some of the latest models were rejecting settings that older ones accepted, which caused parts of the assistant to fail. The assistant now sends each model only the settings it supports.
• Improved dictation UI for Mac. The dictation interface in the Mac app has been redesigned for a cleaner experience. This is an interim improvement, with further polish planned.
• Faster, smoother startup. The assistant could freeze for a few minutes on startup while it initially booted. It now stays responsive the whole time.
• Upgrade reminders for hosted installs. Hosted setups now show a banner when an update is available, without nagging or getting in the way.

A wide release: plugins become a first-class marketplace, your assistant can now pull in a second, more powerful model for hard problems, memory moves to a sharper section-grain v3, and subagents, workflows, Slack, and the Activity page all get upgrades.

• Plugins are now a first-class marketplace. Browse, install, and update plugins from inside the app, alongside the ones the vellum team ships. Plugins are TypeScript packages with lifecycle hooks (distinct from skills, which are instruction bundles), so they can reshape how the assistant behaves, not just what it knows. Read more here.
• Privacy policy and consent updates. We have updated the privacy policy and the Share Analytics / Share Diagnostics control. Share Diagnostics now covers crash reports, conversation traces, and session replay data. Because the consent terms changed, you will be asked to review and reconfirm your privacy-related preferences. Read the updates here.
• Workflows - When your assistant needs to apply the same operation across several things (researching products from different brands, comparing multiple options, summarizing a batch of docs), you can watch each one run in parallel: a progress card in your chat, plus a side panel showing each item start and finish.
• Your assistant can consult a more powerful model. The new Advisor plugin lets the main agent step back on a complex task and consult a second agent running a stronger model before committing to an approach. It comes installed and enabled by default on new assistants. Self-hosted and BYOK workspaces can point the advisor at one of their own profiles.
• Memory v3. Memory moves to a section-grain model: instead of one flat page per topic, your assistant retrieves a compact card plus the specific section that matches, so recall is sharper and less bloated. New assistants get v3 from creation. If you are on an existing assistant, ask it to run the "Memory v3 Migration" skill to reshape your current memory from v2 into the new format. It reads v2 pages either way, but retrieval is noticeably better once migrated.
• Firecrawl joins both web search and web fetch. Your assistant can now use Firecrawl for either searching the web or fetching a specific page, alongside the existing options (Perplexity, Brave, Tavily). Firecrawl returns clean, relevance-ranked results and has a freshness filter.
• Smarter, tidier subagents. The subagent detail panel is faster and cleaner: the timeline is virtualized, the panel no longer re-renders on every streamed token, and side-panel open is snappier. Spawned subagents now inherit the inference profile of the agent that launched them, so parallel work runs on the right model without extra setup.
• Balanced profile now runs on MiniMax M3. The managed Balanced profile is repointed to MiniMax M3 on Fireworks (the former Balanced Economy configuration), keeping the same Balanced label while improving the cost-to-quality tradeoff. The separate Balanced Economy profile has been deprecated.
• Slack improvements. Direct-message replies no longer disappear. Approval cards and reactions clear across every surface when resolved. Access-request verification codes are delivered straight to the requester.
• A cleaner Activity page. The page formerly known as Home is now Activity, decluttered and polished: Notifications and Schedules tabs, a slide-in detail drawer that respects reduced-motion, persistent accordion state, and system schedules (Heartbeat, Memory Consolidation, Memory Retrospective) restored to the view.
• OS-Beta profile for open-weight model testing. A new model profile, OS-Beta, lets you opt into early-access open-weight models. It currently points to GLM 5.2 via Fireworks and is gated behind a feature flag.
• Chrome extension for self-hosted assistants. The extension now works again with self-hosted setups: browser-use is functional, and you can edit the gateway URL or see a clear error when the connection fails.
• Global status banner. A new banner surfaces your assistant's lifecycle state at the top of the app: upgrading, offline, waking up, and any other connectivity events. You'll always know at a glance whether your assistant is actually running.

A new Mac app replaces the legacy Swift build, with feature parity across web and iOS plus stability and performance wins.

• A new Mac app. The Electron-based desktop app replaces the legacy Swift build and keeps feature parity with the web UI and iOS. It is more stable (less prone to app crashes) and snappier across the app.
  • Download the new DMG from vellum.ai/downloads
• Login security improvements. All clients got a security pass: more secure login and token management across the board.

Overall polish everyday feel of the assistant: in-progress steps inside chats show a live timer and the assistant's thinking hovers to reveal how long it reasoned; you can search inside a long chat to jump to a specific word or moment; messages stream more smoothly and stay ordered even when the connection stutters; and a new Balanced economy profile gives you a middle-ground option that keeps costs down without giving up much quality.

• Sign in through your browser. Logging in now hands you off to your normal browser (Safari, Chrome, whatever you use) and brings you back. Small change, but your password manager, saved logins, and two-factor codes all work the way they normally do.
• See how long things take. In-progress steps show a live timer so you know it's working. Hover over its "thinking" to see how long it spent reasoning. Less guessing about whether it's stuck.
• Find within a conversation. Search inside a long chat to jump to a specific word or moment.
• Smoother streaming. Messages hold together more reliably even if your connection stutters. Fewer replies arriving out of order, duplicating, or vanishing.
• A new budget-friendly profile for keeping costs down without giving up much quality. A middle ground between cheapest and best. We call it Balanced economy.

Web searches inside chats show as clean, clickable cards you can verify, adds MiniMax M3 as a new model option, and refreshes the skill lineup with a new geo-audit skill, a rebuilt app builder, and document writing that consistently uses the proper editor. Plus, opening the Mac desktop app now installs the Vellum CLI into your shell PATH automatically.

• MiniMax M3 added as an option.
• Web searches are easier to follow. When the assistant looks something up, results appear as clean, clickable cards in the conversation. See what it searched and click straight to the sources.
• New geo-audit skill for location and SEO analysis, now in the catalog.
• App builder rewrite. The skill that builds dashboards, simple tools, and pages was rebuilt to be more reliable, so what you ask for comes out working more consistently.
• More reliable document writing. Ask for something substantial and it now reliably uses the proper editor, giving cleaner, better-formatted results.
• The Mac desktop app now installs the Vellum CLI into your shell PATH automatically, so running vellum from any Terminal window works without a separate install step. The wrapper points at a locator file that the app refreshes on every launch, keeping vellum message, vellum upgrade, and shell completions available and pointing at the live CLI binary.

We did a minor patch just for this one thing:

Claude Fable 5 is now available as a model option. Only to remove it in a later build due to this: https://www.anthropic.com/news/fable-mythos-access

Pro subscribers bundle credits directly into their plan, makes the daily heartbeat calmer by default, gives scheduling a friendlier builder with plain-language descriptions and a tidier list that hides past one-time tasks and system jobs, and lets you download all your invoices, including Pro and ad-hoc credit purchases, as a single zip.

• Pop conversations into their own windows. Open any conversation in its own window so you can keep one chat visible while working in another. There's also a new "Quick Input" window: a small box to dash off a message without opening the full app or losing your place.
• Keyboard shortcuts. Set your own system-wide hotkeys that work no matter what app you're in. Press it and the assistant pops up, or dictation starts, instantly.
• Native notifications. Alerts now arrive as proper macOS notifications with buttons you can click to act, without opening the app first.
• A cleaner icon and lots of polish. A sharper app icon plus many small visual fixes so it looks at home on macOS.
• Easier scheduling. A friendlier builder, with a plain-language description of when it'll actually run ("every weekday at 9am") so you're not second-guessing the timing.
  • Tidier schedules page. Old one-time tasks are collapsed and system jobs are tucked away, so you see the schedules you care about.
• Calmer by default. The heartbeat now runs at most twice a day out of the box, and you can pause or turn it off whenever. Less noise, more control.
• You can now bundle credits as part of your pro plan: Clearer credit purchasing. The Pro credit screen is cleaner and shows your base monthly fee up front. No surprises.
• Download all invoices at once in a single zip instead of clicking through one at a time. Includes invoices for pro plan and ad hoc credit purchases.

• Cross-machine device pairing: vellum pair mints a device-scoped, revocable token on the assistant host and prints a pairing bundle, vellum connect import sets up a persistent paired assistant on another machine, vellum client --token opens an ephemeral session, and vellum devices and vellum unpair list, revoke, and forget pairings.
• Richer chat surfaces: new work-result, choice, and copy surfaces, plus managed-OAuth connect and approval surfaces so the assistant can hand you structured results and connection prompts inline.
• Pro plan credit bundles: pick a credit bundle during upgrade and change flows, with the current bundle shown on the billing plan card.
• In-chat find: Edit menu now has Find (Cmd+F) wired to in-chat search on both macOS and web.
• Streaming made visible: thinking blocks render live in the chat transcript, a spinning ring wraps your custom avatar while the assistant is working, and the Dock and menu-bar icons reflect the assistant avatar with a live status indicator.
• Skill browsing: skills carry category metadata and SVG icons in their SKILL.md frontmatter, rendered in the web app with dynamic categories.
• Library app polish: apps open fullscreen on web and mobile, with an Edit button in the Library app viewer.
• Cloudflare tunnel provider: new tunnel option for exposing a local assistant.
• Model profile polish: advanced params shown only after a model is chosen, the un-pinned params now read "Default" instead of "Inherit," and advanced-param spacing tightened up.
• Web search sources: clamped sources expand behind a "+N more" pill.
• Connection reliability: SSE reconnects are self-diagnosing with single-cursor resumable replay, and truncated thinking blocks heal on reconnect.

Highlights

• Improved reliability of web release builds, ensuring shared packages are correctly installed before the build process runs
• Fixed a configuration issue that caused web release builds to not properly enable platform mode, resolving potential runtime errors in the web application

Build: 0.8.7 Commit: cfe9507d34 Built at: 2026-06-03 01:43:06 UTC

A model and reliability update: Claude Opus 4.8 joins the hosted catalog, Gemini 3.5 Flash is now available to cloud users after a missing platform rate card was blocking requests, the sleep policy gains 1-hour and 3-hour idle presets, per-assistant sleep thresholds now survive wake cycles, inbound email is fixed on Resend-registered domains, disk-usage alerts fire an owner email, and API key rotation gets a propagation grace period.

Highlights

• Claude Opus 4.8 now available: Anthropic's latest Opus model is live in the hosted catalog and selectable across all surfaces
• Gemini 3.5 Flash on the hosted service: the model was already selectable in the local UI but a missing platform-side rate card returned a 400 for every cloud request; the rate card and billing entry are now wired in so cloud users can select it without hitting a rejection
• Sleep policy 1-hour and 3-hour presets: two shorter idle-sleep durations join the sleep policy control in Settings alongside the existing 1-day, 3-day, and 7-day options so assistants can be tuned to sleep faster without using the full daily minimum
• Idle-sleep thresholds persist across wake cycles: per-assistant sleep cutoff settings were being reset to the system default after each wake event; the configured threshold now survives correctly
• Disk-usage email notifications: when workspace storage crosses the alert threshold the assistant owner now receives an email with remediation links to upgrade storage or ask the assistant to clean up, alongside the existing in-app error notification
• API key rotation grace period: a short overlap window during key rotation keeps the old key valid while the new key propagates, eliminating the authentication gap that could cause brief request failures on self-hosted instances
• Email inbound receiving fixed on Resend domains: newly registered Resend domains had receiving silently disabled because the create-domain API call was missing the nested capabilities object; inbound replies were being dropped with no error surfaced

A billing and reliability release: MiniMax is now an OpenAI-compatible provider, self-hosted assistants route runtime calls directly to their own ingress, billing and resize flows get confirmation steps and clearer copy throughout, email and domain settings gain a live verification chip and confirmation modals, the CLI gains tab management and roadmap CRUD, and the avatar editor learns to generate from a prompt or reset to default.

Highlights

• MiniMax is now available as an OpenAI-compatible provider: configure MiniMax models through the OpenAI-compatible provider flow with primary and fallback endpoint validation, with macOS thinking UI and conversation title generation updated to recognize MiniMax's <think> reasoning tags so they render in the collapsible thinking surface and never bleed into auto-generated titles
• Self-hosted assistants route runtime calls directly to their own ingress: when you connect a self-hosted assistant from the web app, message and event calls go straight to your assistant with the platform actor token attached, instead of being proxied through the cloud platform
• Billing and resize flows are clearer and safer: tier upgrades go through a read-only confirmation step before submitting to Stripe, plan resize triggers a confirmation modal after the Stripe portal step, adjust-plan failures surface inline instead of failing silently, downgrades show resource changes in read-only resource cards with an inline downgrade path, and the Compute & Resources card uses a unified tile layout with contextual resize CTAs
• Email and domain settings are easier to work with: domain settings show a live verification status chip, domain registration and deletion go through confirmation modals, the managed email upsell auto-opens the upgrade modal, the domain status CLI is wired to the live verification endpoint, and the assistant handle auto-locks when it matches the subdomain
• Disk pressure and credits banners surface action: disk pressure banners are redesigned with clearer state, and the credits exhausted banner now links directly to auto top-ups
• Multi-tab management from the CLI: list, open, switch, and close conversation tabs with new assistant tabs subcommands, with each connected client tracking its own tab scope independently
• Roadmap from the CLI: the assistant roadmap command now supports create, search, and upvote alongside listing, so you can manage roadmap items end-to-end without leaving the terminal
• Avatar editor gains Generate with AI and Reset to default options, so you can spin up an avatar from a prompt or revert to the original without leaving the editor
• Composer emoji aliases: :huff and :steam now resolve to 😤 alongside the existing alias set
• Chat reliability fixes: the thinking indicator no longer survives stale terminal events because send state is now scoped to the active turn, and macOS conversation list pagination keeps the scheduled section from being truncated on large histories
• Share Feedback dialog polish: cleaner layout, design library components throughout, and contextual tips, so reporting an issue feels less like filling out a form
• LaunchDarkly references removed from the repo, with the assistant fully moved to push-driven feature flag invalidation via SSE so flag changes take effect immediately instead of waiting on a polling loop

An identity and model-catalog release: every assistant gets a unique handle that doubles as its subdomain, the document editor gains agent-driven comments and find-and-replace, the model picker adds Gemini 3.5 Flash, Gemini 3.1 Flash-Lite GA, Grok 4.3, and Owl Alpha, OpenAI-compatible providers can finally point at any endpoint on macOS, and a new /clean slash command and a dock-icon fix tighten the day-to-day feel.

Highlights

• Unique assistant handles and matching subdomains: every assistant now has a unique handle that doubles as its subdomain by default, the handle is editable inline from About Assistant, the subdomain locks to match it, and existing assistants were backfilled so handles are non-nullable across the system
• Document editor agent tools: the assistant can now leave inline comments on documents and use grep and sed-style find-and-replace tools, so iterative edits run as tool calls instead of full rewrites
• Custom email address builder in Pro onboarding: the Pro onboarding domain step now includes an inline email-address builder so you can configure a custom email domain end to end during signup, and the new assistant domain register --email-username flag does the same from the CLI
• Expanded model catalog: Gemini 3.5 Flash and Gemini 3.1 Flash-Lite GA via Google, Grok 4.3 and Owl Alpha via OpenRouter, and Fireworks now respects per-model reasoning_effort ceilings so deep-thinking models stop getting clipped by a global cap
• OpenAI-compatible providers on macOS: the provider connection editor now exposes Base URL and Models fields so you can point macOS at any OpenAI-compatible endpoint (vLLM, Ollama, LM Studio, llama.cpp) without leaving Settings
• New /clean slash command: strips runtime injections from the conversation without summarizing the rest of the turn, available in the web chat composer slash picker and on macOS
• Inline file attachments in assistant message bubbles on macOS: file chips now render between paragraphs in assistant replies instead of all batching at the bottom of the message
• macOS performance and reliability: dock icon image processing moves off the main thread to eliminate 2s+ AppHang events, self-signed certificate creation no longer duplicates on every build, and the home feed preserves read-row text contrast on dark backgrounds

A home feed and notifications intelligence release: the macOS home page now separates urgent from routine activity, the CLI gains full schedule management, memory routing is on by default, Fireworks joins the managed provider roster, and the web platform recognizes self-hosted assistants.

Highlights

• Home feed redesigned with urgency routing: urgent items go to Inbox, routine items to Activity, every row shows timestamps and unread weight, read rows fade back, and the Activity section collapses to keep your Inbox focused
• Notification urgency is now first-class: urgent signals get OS banners and Inbox placement, routine signals route to Activity silently, and assistant notifications send --urgent lets your own scripts control urgency
• Full schedule management from the CLI: create, cancel, enable, disable, and delete scheduled tasks with new assistant schedules subcommands, complete CRUD without opening the app
• Memory v2 router is on by default: the Sonnet-based memory router now activates for managed Anthropic installs with a balanced profile, no feature flag needed, retrospective timestamps reflect your local timezone, and transcripts use participant display names
• Fireworks is now a managed connection: use any Fireworks model without supplying your own API key, the same way managed Anthropic, OpenAI, and Gemini connections work
• Self-hosted assistants are now recognized in the web app: if you are running the open-source daemon locally, the platform shows a dedicated screen with a link to settings instead of attempting to start a cloud instance

A cross-platform polish release: subagents get a cleaner design and iOS gets a critical ChatComposer reliability fix.

Highlights

• Subagent UI redesigned to match macOS: cleaner layout, full-screen detail panel on mobile, and consistent design language across platforms
• iOS ChatComposer reliability fix: resolved an infinite update loop triggered by WKWebView re-dispatched input events
• Usage page redesigned with clearer information hierarchy, descending-sorted stacked-bar tooltips, and improved aesthetics

A focused release on a redesigned model profiles and provider connections system, Twilio phone calling for platform-hosted assistants, one-click Publish to Vercel from the app builder, iOS push notifications, a smarter memory pass, and a tightened iOS login flow.

Highlights

• Redesigned model profiles and provider connections: mix managed and bring-your-own connections per profile, edit labels and status inline, save tweaks as new variants, and switch profiles mid-conversation with the new /model slash command
• Twilio phone calling for platform-hosted assistants: your assistant can now answer and place real phone calls when running on the Vellum platform, so you can reach it from anywhere without opening an app
• Publish to Vercel from the app builder: when your assistant builds you an app, it can now deploy it to Vercel for you in one step, with complex apps handed off to the assistant for end-to-end deployment and live URLs surfaced in the deploy toast
• iOS push notifications are live: with the latest staging iOS app, your assistant can send native APNs notifications to your phone for activity completions, scheduled wakeups, and turn-end events you'd otherwise miss on mobile
• Memory v2.5 under the hood: a background retrospective pass reflects on recent conversations, a Sonnet-based router replaces dynamic context injection for memory recall, and per-conversation memory state forks cleanly with the conversation. No visible UX changes, just memory that feels right more often
• iOS login flow polish: a single Sign In button on the Capacitor login screen, native OAuth completion now lands through a deep-link router instead of falling back to popups, conversational OAuth opens in SFSafariViewController, invalid_grant errors prompt a reconnect instead of silently failing, and the onboarding name is pre-filled from your authenticated profile
• Tavily joins the web search provider list, giving you another option alongside the existing providers
• Quality-of-life polish: a Discord community card and chat banner make it easier to find the Vellum community, assistant changelog brings release notes into the CLI, NOW.md injection is now configurable, the macOS provider editor gained a Save as New button, Google OAuth setup is now single-step with inline credential collection, and gateway logs are pretty-printed with a JSONL sidecar for log-tail

A major intelligence and cross-platform release: Memory v2 is now live, the web app can reach back to your Mac for local actions, managed integrations expanded, and the iOS experience is smoother across login, settings, attachments, and notifications.

Highlights

• Memory v2 is live: your assistant now builds richer concept pages instead of isolated fact snippets, retrieves the right context more reliably, consolidates what it learns in the background, and gives you cleaner memory surfaces for browsing and searching what it knows
• Use your Mac from the web app: when your desktop assistant is connected, web conversations can route local actions back through your Mac for workspace files, terminal commands, browser control, and other host-side work without forcing you to switch clients
• Live integrations now include Discord, GitHub, Google, HubSpot, Linear, Notion, Twitter, Asana, Outlook / Microsoft, and Todoist, with smoother native OAuth handoff on iOS and desktop
• Model settings now include Fireworks and OpenRouter in the provider catalog, making it easier to configure assistants that use your preferred model provider
• iOS got a serious polish pass: Sign in with Apple, native OAuth completion, APNs push notification groundwork, better mobile settings layouts, improved attachment preview and download behavior, file downloads through the native share sheet, and cleaner login screens
• Everyday reliability improved across notifications, trust-rule chips, document persistence, skill loading, gateway routing, OAuth setup, memory maintenance, and macOS signing and update flows

Highlights

• Improved memory retrieval routing for more reliable and efficient background memory lookups
• Under-the-hood stability improvements to how the assistant accesses and recalls stored information
• General reliability fixes to the memory system to reduce retrieval inconsistencies

A broad reliability release focused on the Chrome extension and cloud connection path, macOS workflow polish, contact and communication flows, heartbeat and schedule reliability, and tighter gateway security.

Highlights

• Chrome extension and cloud reliability improved across the board: requests now send fresh session tokens consistently, platform API calls use the right headers and base URLs, self-hosted pairing is fixed, and the extension is prioritized over the macOS SSE bridge for browser control
• Gateway and security hardening tightened local and cloud boundaries, including stricter loopback handling, known-origin checks for pairing and CORS, safer token refresh behavior, secret redaction before recall evidence reaches the model, and reduced elevated capabilities in Docker-mode assistant containers
• Contact, messaging, and document workflows got smoother with contact prompt commands and macOS panels, gateway-owned Slack contact upserts, missed @mention catch-up after socket reconnects, stronger trusted contact checks, and PDF export for Document Writer
• Conversation workflows are more resilient with reliable notification deep links, Mark as read in the conversation context menu, a cleaner conversation-switch loading skeleton, visible trust-rule save errors, and the new burst-based Worked for X.Ys activity model
• Open-source visibility improved in the macOS app with a new Settings card for the public GitHub repo and a View on GitHub link in the About panel
• Scheduling, heartbeat, and operations are more robust: scheduled tasks can retry after failures, heartbeat runs use cron-style timing with missed-run detection, SSE disconnects are detected faster with a heartbeat watchdog, and assistant gateway logs tail makes gateway debugging easier

A focused polish release with stronger Language Model controls, more reliable trust rule editing, Chrome extension fixes, macOS quality-of-life improvements, and credential infrastructure upgrades for smoother local and cloud assistant setups.

Highlights

• Language Model controls got more precise: profile-specific context budgets, model-aware max token sliders, refreshed context metadata, and effective context handling across the main agent loop, wake paths, and slash commands make model profiles behave more predictably
• Your Own mode is more reliable: explicit user saves are now respected, main agent profiles can override static defaults, profile context and slider fallbacks are preserved, and actual provider metadata is stamped correctly when routing picks a non-default provider
• Trust rules and approvals are easier to understand: rule editor copy now shows natural language instead of structured tool data for non-bash tools, trust badge taps open the right edit flow, and approval provenance is tracked with mode, reason, and risk threshold fields
• Chrome extension reliability improved with fixes for status probing, activity isolation, privileged tabs, popup icon paths, self-hosted UX, and hostname handling across local and cloud modes
• macOS polish and workflow fixes: turn-end notifications now fire when the app is unfocused, task progress widgets default inline with a pop-out option, the profile editor is cleaner, toolbar state is better isolated from conversation list updates, and account deletion requests can be started from the app
• Credential and CLI infrastructure moved forward with credential account key normalization, a direct CES credential management CLI, API key migration work, and a new vellum upgrade --latest flag for pulling the newest available version

A redesigned trust and permissions system with a new v3 rules engine, Gemini 3 model support, GPT-5.5 as the new default OpenAI model, and a wide range of new CLI commands and a reworked browser extension experience.

Highlights

• Redesigned trust and permissions system: a new v3 Trust Rules engine replaces the older permissions model, with cleaner presets (Conservative, Relaxed, Autonomous), suggested rules with an Allow & Create Rule button, directory-scoped rules, and a fully redesigned Trust Rules management UI
• Gemini 3 model support and catalog improvements: Gemini 3 models are now available with correct pricing, thought signature capture, and tool-call metadata, the default OpenAI model is now GPT-5.5, and OpenAI reasoning effort can now be explicitly disabled or set to an extra-high tier
• Expanded CLI: new commands for inspecting installed skills, managing trust rules, registering and listing webhooks, setting and getting environment variables, verbose exec, and SSH and exec support for managed instances
• Reworked browser extension: a proper onboarding flow, SSE-based event transport, and direct gateway pairing replacing the legacy native messaging host, for a smoother and more reliable connection between your assistant and Chrome

The Vellum browser extension is now one-click install from the Chrome Web Store, X/Twitter is generally available as a managed integration, and Slack setup is one click with expanded permissions so your assistant can triage your messages, not just post. Plus a new Voice settings tab on the web platform, a conversation Refresh action, and a wide range of macOS stability and polish fixes.

Highlights

• Vellum browser extension on the Chrome Web Store: install the Vellum Assistant Browser Relay from the Chrome Web Store in one click, no more loading an unpacked extension in developer mode. The extension bridges your assistant to your live Chrome tabs for reading, clicking, filling, and extracting on any site you are already signed into, and now also supports cloud sign-in directly from the extension
• X/Twitter integration is generally available: connect X to your assistant through the managed integration without toggling a flag, with expanded scopes so your assistant can read likes and bookmarks in addition to posting and browsing
• Slack is more capable and easier to set up: creating a Slack app is now a single step with the redirect URL baked into the manifest, and the install flow has been simplified so broken links during app creation are gone. Permissions were also expanded so your assistant can now read your Slack messages for triage and summarization in addition to posting, and the bot automatically joins public channels instead of requiring an invite
• Voice settings on the web platform: the web platform now has a dedicated Voice settings tab matching the macOS app, with push-to-talk presets, and the Test button in voice settings is always enabled so you can preview voices at any time
• Conversation Refresh action: a new Refresh action in the conversation menu lets you reload a conversation when its state gets out of sync with the server
• macOS stability and polish: dedicated network session for background streaming to fix stalled responses, cold-launch avatar caching, fixed scroll white space and push-to-top jitter during streaming, sidebar skeleton that waits for restoration instead of a timer, smoother home panel transitions, and a wide range of reliability improvements across memory compaction, onboarding, credentials, and OAuth

Claude Opus 4.7 support, a migration to OpenAI's newer Responses API, major macOS performance fixes, a smoother chat scroll experience, configurable log retention, improved Gmail cleanup, and broad stability polish.

Highlights

• Claude Opus 4.7 support: the newest Anthropic model is now available across the app and is automatically used when your assistant needs its strongest quality-oriented reasoning
• OpenAI provider moved to the Responses API: upgraded the underlying connection to OpenAI models for better streaming, tool calls, and compatibility with newer features
• Major macOS performance improvements: resolved multiple two-second-plus app hangs caused by layout, font, avatar, sound, and menu-bar initialization, plus a new inverted scroll architecture for noticeably smoother chat scrolling
• Configurable LLM log retention: choose how long request logs are kept on your device from Settings > Permissions & Privacy. Options are 1, 7, 30, or 90 days, or never expire
• Faster, more accurate voice transcription: Google Gemini speech-to-text now streams over the Live API for real-time partial transcripts, with support for speaker labels when the provider offers them
• Smarter Gmail cleanup: persistent blocklist and safelist preferences, a new cold-outreach workflow with automatic classification and enrichment, and more reliable archiving behavior
• Conversation archive: archiving and unarchiving conversations now syncs reliably with the server, with archived items sorted by when they were archived
• Assistant thinking in progress cards: the progress card now includes a thinking sub-row and factors thinking time into the total duration, so you can see where the assistant is spending its time
• Polish and stability fixes: cleaner thinking block layout within the chat column, Reflections grouped under Background in the sidebar, wider model names in the Usage breakdown, a fixed Web Search API key field, smoother onboarding, and a wide range of reliability improvements across memory, credentials, and OAuth

Dramatically improved chat performance and stability on macOS, a redesigned onboarding flow, real assistant names throughout the UI, a refreshed integrations panel, and broad UX polish.

Highlights

• Dramatically improved chat performance and stability on macOS: resolved a wide range of rendering hangs, layout freezes, and scroll issues, including fixes for blank chat on conversation switch, streaming scroll corrections, and smoother message rendering overall
• Redesigned onboarding flow: the setup experience now focuses on what you want your assistant to do and how it should behave, replacing the previous personality quiz with a more practical, goal-oriented approach
• Assistant identity improvements: assistants now display real names instead of IDs throughout the app, with random name generation for new assistants and the ability to edit role and description directly from the Identity panel
• Refreshed integrations panel: the integrations page has been redesigned with a cleaner grid layout and moved into Settings as a dedicated tab, with provider logos for easier recognition
• Smoother scrolling and chat polish: a new scroll model brings smoother conversation scrolling, better alignment throughout the chat, animated conversation names on hover, a new sidebar groups divider, and archived conversations sorted by archive time
• Memory and context improvements: memory recall is now more configurable, consolidation is less aggressive to prevent memory loss, and deleted memories are now recoverable
• Broad stability and UX fixes: simplified permission controls, skill previews now work before installation, Slack gained /new command support, and numerous fixes across modals, attachments, sidebars, and the composer

Introduces a referral program for earning credits, Linear integration, cleaner thinking blocks, configurable knowledge base injection, improved API key management, a polished chat and sidebar experience, and broad stability fixes.

Highlights

• Referral program: invite friends with your unique referral link and earn credits when they sign up. Accessible from the Billing tab and the sidebar, with stats tracking and program details built in
• Linear integration: connect to Linear through the platform without configuring your own OAuth app, enabled by default during onboarding alongside Outlook
• Thinking blocks rendered as markdown and collapsed by default: the assistant's reasoning is now displayed as clean, formatted text that starts collapsed, making responses easier to read
• Configurable knowledge base injection: control which knowledge base files are automatically included in conversations, with fixes for duplicate and out-of-order content
• Improved API key management: a more reliable system for storing and reading API keys across settings and services
• Chat and sidebar polish: a new animated typing indicator, spell checking in the composer, a Recents group replacing ungrouped conversations, and cleaner sidebar text handling
• Stability and correctness fixes: numerous fixes across scrolling, history loading, dictation, audio handling, and OAuth flows, reducing crashes and stale data issues

Introduces the Personal Knowledge Base for reliable fact recall, a refreshed design system, major macOS chat performance fixes, background agents for parallel task execution, and numerous stability improvements.

Highlights

• Personal Knowledge Base (PKB) introduced: your assistant now has a persistent memory system that files and retrieves important information across conversations, with a fully redesigned Memory inspector showing what your assistant remembers and how confident it is
• Refreshed design system: updated color palette, new typography, and polished components throughout the app for a cleaner, more modern look
• Major chat performance improvements: resolved several severe slowdowns on macOS, including multi-minute freezes when switching conversations and layout bottlenecks that caused the app to hang
• Background agents for parallel work: your assistant can now spawn background agents to handle tasks independently, with different roles for different types of work like research, coding, or planning, and the ability to report results back when finished
• Stability and rendering fixes: smoother thinking block animations, improved markdown formatting, better sidebar behavior, and more reliable scrolling throughout the app

The biggest release yet: Vellum goes open source, introduces platform-hosted assistants, a completely revamped memory system with image support, Outlook feature parity, conversation folders, and proactive assistant check-ins.

Highlights

• Open source launch: the Vellum Assistant repository is now publicly available, inviting the community to explore, contribute, and build on the platform
• Platform-hosted assistants: assistants can now run fully hosted on the Vellum platform, removing the need for local infrastructure and enabling seamless cloud-based operation
• Revamped memory system with image support: the memory system has been completely rebuilt with support for image references, smarter search that combines multiple retrieval strategies, and more reliable memory consolidation
• Outlook Calendar and Email reach full feature parity with Google: Outlook Calendar and Outlook Email integrations are now generally available, matching the functionality previously available only for Google Calendar and Gmail
• Conversation folders and sidebar improvements: organize conversations into folders with automatic grouping by source, improved icons and count badges, and easy group management
• Proactive assistant check-ins: your assistant can now periodically review its notes, reflect on recent conversations, and reach out when it has something worth sharing, enabled by default
• Skills system redesign: skills have been rebuilt with better discovery, easier installation, and more reliable behavior across the board
• Performance and stability improvements across macOS and iOS: faster conversation loading, smoother scrolling, and reduced memory usage through better caching and background processing

Major macOS performance and stability improvements, Outlook messaging support, smarter assistant memory, security hardening, and polished UI components.

Highlights

• Significant macOS performance and stability improvements: fixes for chat scroll freezes, sidebar lag, and app responsiveness issues, resulting in a noticeably smoother experience
• Outlook messaging support: Vellum can now connect to Microsoft Outlook as a messaging provider, joining the existing Slack integration and expanding where your assistant can be reached
• Smarter assistant memory: the assistant now remembers its capabilities and available tools at startup, with improved search for finding relevant memories
• Security hardening: removed the ability to bypass permission prompts, tightened access controls across the app, and added stricter validation for sensitive operations
• Polished UI components: redesigned skill detail page, improved file browser, better dropdowns and navigation, and a context window indicator showing how much conversation space is left

Signing key handling improvements and automatic migration for smoother upgrades.

Highlights

• Improved signing key handling to prevent potential authentication issues during normal operation
• Automatic key migration when upgrading from older versions, ensuring a smooth upgrade experience without manual steps

Thinking blocks in chat, overhauled memory and retrieval, /compact command, expanded model support, and collapsible sidebar sections.

Highlights

• Thinking blocks are now visible in chat: see your assistant's reasoning process inline as collapsible blocks, giving you transparency into how responses are formed, enabled by default
• Significantly improved memory and retrieval: smarter memory extraction, better search diversity to surface unexpectedly relevant memories, and improved formatting of recalled information
• New /compact command and context window indicator: manually trigger conversation compaction at any time, with a color-coded bar in the toolbar showing how much context space is remaining
• Expanded model support: DeepSeek, Qwen, Mistral, Meta, Moonshot, and Amazon models added through OpenRouter; Anthropic 1M context window beta and fast mode now supported
• Collapsible sidebar sections and channel conversations: Scheduled and Background sidebar sections can now be collapsed, and channel-bound conversations display with a read-only indicator

Build: 0.5.13 Commit: ac3ecb51c Built at: 2026-03-27 16:03:20 UTC

Build: 0.5.12 Commit: 5dda41f46 Built at: 2026-03-26 22:27:05 UTC

Build: 0.5.11 Commit: 68f3317b5 Built at: 2026-03-25 17:55:35 UTC

Build: 0.5.10 Commit: f96060eaf Built at: 2026-03-24 15:19:58 UTC

Build: 0.5.9 Commit: df5d471b1 Built at: 2026-03-24 04:22:15 UTC

Build: 0.5.8 Commit: 42106414e Built at: 2026-03-24 01:54:38 UTC

Build: 0.5.7 Commit: fba21649a Built at: 2026-03-23 15:06:56 UTC

Build: 0.5.6 Commit: 8ce415d9b Built at: 2026-03-20 13:47:26 UTC

Build: 0.5.5 Commit: bef29f63e Built at: 2026-03-20 00:59:07 UTC

Build: 0.4.55 Commit: 1e420a4 Built at: 2026-03-14 18:13:00 UTC

Build: 0.4.53 Commit: e8cb485 Built at: 2026-03-14 03:37:14 UTC

Build: 0.4.52 Commit: 580d560 Built at: 2026-03-13 21:28:12 UTC

Build: 0.4.51 Commit: 584420c Built at: 2026-03-13 21:07:56 UTC

Build: 0.4.50 Commit: ac0845f Built at: 2026-03-13 16:32:29 UTC

Build: 0.4.49 Commit: 9410d83 Built at: 2026-03-12 17:55:52 UTC

Build: 0.4.48 Commit: 9a98afb Built at: 2026-03-12 00:05:31 UTC

Build: 0.4.46 Commit: 8bdec19 Built at: 2026-03-11 17:56:24 UTC

Build: 0.4.45 Commit: c23f8bc Built at: 2026-03-11 13:05:26 UTC

Build: 0.4.44 Commit: 75549ac Built at: 2026-03-10 00:24:03 UTC

Build: 0.4.43 Commit: c2437a1 Built at: 2026-03-09 16:38:27 UTC

Build: 0.4.42 Commit: a333764 Built at: 2026-03-06 21:49:11 UTC

Build: 0.4.41 Commit: ad0eb00 Built at: 2026-03-06 21:03:52 UTC

Build: 0.4.40 Commit: 39acbb1 Built at: 2026-03-06 20:19:51 UTC

Build: 0.4.37 Commit: f98bad6 Built at: 2026-03-06 12:25:18 UTC

Build: 0.4.36 Commit: bf07d08 Built at: 2026-03-06 11:25:53 UTC

Build: 0.4.35 Commit: cbaf75e Built at: 2026-03-05 21:21:12 UTC

Build: 0.4.34 Commit: c042e90 Built at: 2026-03-05 20:24:03 UTC

Build: 0.4.33 Commit: 413e80b Built at: 2026-03-05 18:07:24 UTC

Build: 0.4.32 Commit: 82bb915 Built at: 2026-03-05 14:00:56 UTC

Build: 0.4.30 Commit: b83460b Built at: 2026-03-05 01:10:15 UTC