A personal AI assistant that lives on your computer. It can take real actions on your behalf: reading files, sending emails, browsing the web, controlling your Mac, building apps, managing your schedule, making phone calls, and more. It has its own identity, personality, and long-term memory that persists across conversations. See What is Vellum? for the full overview.
Those are conversation tools. You type, they respond, you copy-paste the answer somewhere. Vellum is different because it has tools (it can actually do things on your computer and across services), memory (it remembers you across conversations with a full hybrid search system), a persistent identity (its own personality, name, and behavioral rules you can customize), and it reaches you everywhere (desktop, Telegram, Slack, voice calls). More detail in What is Vellum?.
You can run Vellum locally with your own Anthropic API key at no cost beyond your API usage. Vellum also offers a managed mode where you sign in with a Vellum account and the assistant runs on our platform.
On Vellum Cloud, you can reach your assistant from any modern browser at vellum.ai, the iPhone and iPad app, the macOS desktop app, and a command-line interface. Beyond those first-party surfaces, channels include Telegram, Slack, email, and phone calls.
Three paths, picked by what you care about most:
Most users should start on Cloud. You can move later if your needs change. See the Hosting options overview for the full comparison.
Yes. There's a native iPhone and iPad app on the App Store, and you can also reach your assistant through Telegram or phone calls from any device.
On Vellum Cloud (the default), your workspace, memories, and credentials live in your private, encrypted cloud account, not shared with other users. If you self-host, all of that stays on your machine instead, with credentials kept in the macOS Keychain or an AES-256-GCM encrypted file, isolated behind a separate Credential Execution Service. In both cases, your conversations and context are sent to the AI model provider (Anthropic) to generate responses. That's the trade-off we're transparent about. Full details in Privacy & Data.
No. From our side, your data is not used for training or fine-tuning. Anthropic's API terms also state that API data is not used for model training. We recommend reading Anthropic's Privacy Policy directly for the most current details.
Only if you opt in. There are two optional toggles in Settings > Privacy: usage analytics (anonymized token counts and feature adoption — no message content) and crash diagnostics (error reports via Sentry — no personal data). Both are off by default.
Your assistant is yours. Vellum doesn't have a dashboard, an admin panel, or any way for your employer to see your usage. That said, if you're using your employer's computer or network, they could potentially see the API calls or web traffic from your assistant. If that's a concern, self-hosting on a personal device is the most private option. Use your judgment based on your work environment.
On Vellum Cloud, your account and workspace stick around unless you ask us to delete them. You can export your workspace at any time, or request full deletion through the account settings or support. If you self-host, your workspace folder (~/.vellum/) stays on your machine until you delete it manually; vellum retire from the CLI archives the workspace as a tarball before removal. See Security Best Practices for the full reset process.
Yes, in a controlled way. You can grant trusted contacts limited access to your assistant through channels like Telegram or Slack. Trusted contacts can chat with your assistant but can't access your memories, modify your workspace, or use sensitive tools without your explicit approval. Unverified people who message your assistant get heavily restricted access. See The Permissions Model for details on how trust gating works.
A lot. Gmail management, Google Calendar, Slack integration, web browsing, computer control, phone calls, image generation, coding, app building, document writing, task management, screen watching, media processing, and more — about 30 built-in skills in total. You can also build custom skills to extend it further. See the Skills Reference for details on each capability.
Files inside the workspace (~/.vellum/workspace/) are accessible without prompts. Files outside the workspace — on your host machine — require your explicit permission each time. You see what file it wants to access, whether it's a read or write, and can choose to allow once, allow temporarily, or create a persistent rule. See The Permissions Model.
With the Gmail skill, your assistant can draft and send emails from your Gmail account, but sending always requires your explicit approval. It creates a draft first, and you approve before anything is sent. Your assistant can also use its own email address through AgentMail for sending on its own behalf.
Yes, with your permission. The Computer Use skill lets your assistant see your screen (via accessibility APIs and screenshots) and control mouse and keyboard input. This requires macOS Accessibility and Screen Recording permissions, and each action is prompted individually for approval. Sessions are capped at 50 steps with loop detection and destructive action blocking. See Computer Use.
Yes, in three ways. You can tell it facts and preferences naturally in conversation (it extracts and saves them to long-term memory automatically). You can edit its workspace files directly (SOUL.md for behavior, USER.md for facts about you). And you can build custom skills that give it entirely new tools and capabilities. See Tools & Skills.
No. Your workspace and tools are local, but your assistant needs an internet connection to think — it sends your messages to the AI model provider (Anthropic) to generate responses. Without internet, it can't respond.
Yes. Say “I want to rename you to [name]” or edit IDENTITY.md directly in ~/.vellum/workspace/.
Yes. Tell it what you want (“be more casual,” “stop being sarcastic”) and it will update SOUL.md. You can also edit SOUL.md directly. Changes to workspace files take effect on the next conversation. See The Workspace.
Yes. Your assistant has a customizable avatar. Say “put on a wizard hat” or “change your color to emerald” and it will update its appearance. Avatar customization is managed through the assistant's identity and style settings.
Anthropic's Claude by default. The model can be changed in config.json in your workspace.
No. It's very good at sounding like it does, though. Don't let the personality fool you. It's a language model with a good costume.