A personal AI assistant for files, connected services, and recurring tasks, with editable memory and self-hosting. Its hosted Personal Assistant Platform closes on November 2, 2026.
Used this tool? Rate it
Used this tool? Rate it
Vellum is a personal AI assistant that combines conversations, persistent memory, and tools for working across files and connected services. It can read files, manage calendars, browse, run code, and act on tasks.
The hosted Personal Assistant Platform is scheduled to close on November 2, 2026. Vellum says cloud-hosted personal assistants will stop functioning, website login will become unavailable, and their data will be permanently deleted. This notice was visible on October 27, 2026; the hosting, pricing, and capability descriptions below sit within that announced lifecycle.
Separate developer documentation still describes low-code AI workflows that chain model calls and business logic, deploy behind an API, and retain versioned definitions and per-step invocation logs. It also describes evaluation Test Suites and Metrics scoring between 0 and 1. Those descriptions do not establish entitlements of the current assistant plans.
The shutdown notice names the Personal Assistant Platform and its closed-source cloud platform. It does not explicitly resolve the commercial status of every older workflow, evaluation, API, or SDK offering.
Vellum's persistent memory organizes knowledge into linked concept articles and retrieves relevant parts for a conversation. Facts, preferences, corrections, plans, and relationships can survive across sessions. The system distinguishes this stored knowledge from the immediate context available to a model, which can contain recent messages, selected memories, workspace files, and tool results.
Recall is selective: the assistant receives a bounded set of relevant memories, and no individual memory is guaranteed to appear on every turn. It is not intended to preserve every sentence as long-term knowledge.
The assistant's workspace uses editable text files for its identity, behavior, and knowledge about the user. In cloud hosting these are browsed through the web app; local hosting exposes the corresponding folder on the user's machine. Credentials occupy a separate protected store and are excluded from workspace exports and diagnostic logs.
Skills are loaded when their descriptions and activation hints match the conversation. The instructions do not run by themselves.
Model profiles set the workspace default for conversations, background memory work, and scheduled tasks. A conversation can use its own profile, and action overrides can assign profiles to individual action types. Custom profiles can select providers or models beyond the built-in choices. Switching a conversation's profile leaves the workspace default in place.
Teleport offers a desktop Move to Local flow under Preferences → Settings → General. Vellum warns that it is not fully battle tested and may fail for some assistants. Its documented process includes logging out and likely supplying independent model keys and replacement services.
For hosted-data export, a signed-in user selects the assistant at the export interface and chooses Export and download. The assistant pauses in Recovery Mode while the archive is packed, then can be resumed. The export file is removed after seven days; that file lifetime differs from the platform's shutdown deletion date.
Vellum supports both one-time schedules and recurring tasks described in natural language. A one-time task has a firing time and is disabled after it fires. Recurring schedules use cron or RRULE patterns underneath, including calendar-style exceptions and bounded recurrence. Documented examples include a morning summary of calendar events and unread email, or a weekly update assembled from Slack activity.
Watchers poll supported services and apply an action prompt when new events arrive. The documented providers include Gmail, Outlook, GitHub, Google Calendar, Outlook Calendar, and Linear.
Playbooks handle incoming-message patterns using a trigger, action, channel, priority, and autonomy level. They can execute automatically, prepare a draft, or notify the user. For example, an incoming meeting request can lead to a draft reply rather than an immediate send.
OAuth integrations expose service-specific tools after authorization, and access can be revoked. One Google connection covers Gmail, Calendar, Drive, Docs, Sheets, Slides, and Contacts. Notion access is limited to pages to which its integration has been added.
Vellum positions the current assistant as a personal resource rather than a shared team tool. The central use is one person's identity, preferences, memory, and connected accounts.
User-hosted remote operation is documented for technical users who manage infrastructure, credentials, networking, and security isolation. It can provide an always-on runtime without requiring a personal laptop to remain awake.
The notice explicitly preserves the case of assistants originally configured without a Vellum account using independent keys and OAuth apps.
The downloads page lists the following platform requirements.
| Surface | Listed minimum | Role |
|---|---|---|
| Mac | macOS 15 Sequoia | Native desktop application with computer use |
| Windows | Windows 10 | Desktop application for Windows PCs |
| iPhone | iOS 15 | Mobile access to the assistant |
| Android | Android 7.0 | Mobile access to the assistant |
| Chrome extension | Chrome 120 | Browser automation extension |
The web channel is cloud-only and cannot access local host files, run host shell commands, control the computer, or watch the local screen.
Desktop supports computer use and host-file or shell access on Mac and Windows. Some interface features differ: the Fn voice key and floating companion are macOS-only, while Windows uses a configurable voice shortcut. Windows computer control works in normal, non-elevated windows.
Native local hosting runs the assistant on a Mac or Windows PC and keeps its workspace there. It still sends prompts and relevant context to the configured AI provider unless a local inference arrangement is used. Local availability depends on the computer staying awake. The hosting documentation marks Docker and Apple Container paths as coming soon, so those labels do not establish released local container options.
The current assistant developer guide describes an authenticated server-sent event stream for assistant output and SSH port forwarding for remote access. Stream events include incremental text, tool activity, completed messages, and approval requests.
As captured on October 27, 2026, the assistant pricing documentation lists the following monthly packages.
| Pro package | Monthly price | Machine | Storage | Included monthly usage | Platform fee, email and subdomain |
|---|---|---|---|---|---|
| Mighty | $30 | Small: 1 vCPU, 2 GiB RAM | 10 GiB | Mighty Usage | Not included |
| Super | $100 | Medium: 2.5 vCPU, 5 GiB RAM | 30 GiB | Super Usage | Included |
| Ultra | $200 | Large: 4 vCPU, 8 GiB RAM | 60 GiB | Ultra Usage | Included |
Base is described as free, with a Small machine, 6 GiB of storage, managed model credentials, and pay-as-you-go credits without a monthly minimum. The published free description does not override the announced end of the hosted platform.
Included usage is spent before purchased credits, resets each billing cycle, and does not roll over. Purchased pay-as-you-go credits have a twelve-month expiry. Work that consumes usage includes model inference, web search, image generation, and paid third-party APIs reached through managed OAuth. The usage view attributes spending by day, model, and action; actions requiring paid usage pause when the balance is exhausted.
The terms denominate one Vellum Credit as $1.00 USD and say credits are non-refundable except where applicable law requires otherwise. This credit rule is separate from the shutdown-specific subscription refund: active paid subscriptions on November 2 are automatically cancelled with a prorated refund for the unused billing period.
The main pricing FAQ says upgrades apply immediately, computer and storage changes are prorated, and downgrades or cancellations take effect at the end of the current billing period. The detailed pricing documentation gives a different timing for a smaller machine: immediate application with a credit toward the next invoice. It says smaller usage bundles start at renewal and there are no mid-cycle cash refunds. These published timing descriptions are not identical.
The general trial terms define expiry by the trial period specified in the service, allow one free trial, and permit automatic continuation and charges if the agreement is not terminated by its last day. They do not supply a universal number of trial days.
OpenClaw and Hermes Agent address the personal-agent task more directly than a comparison limited to older prompt-evaluation platforms.
The assistant harness remains self-hostable, but Vellum says it will stop actively maintaining the repository and regularly publishing releases. The cloud platform is closed source. Continued local execution therefore does not carry a promise of continuing vendor maintenance, hosted account access, or managed credentials.
In the pages reviewed here, the documented export-file lifetime is a specific file-removal rule, not a general promise about log or conversation-history retention.
Private conversations have a separate memory scope. They can read from shared memory, but newly saved memories remain scoped to the private conversation. That controls memory reuse; it does not mean the conversation bypasses the selected AI provider or ceases to require a model call.
The privacy notice allows collection of conversational content consistently with user preferences, including sent messages, files, and generated responses, for service evaluation and improvement. It says Vellum does not use that content to train AI models and provides opt-in or opt-out controls during onboarding and in privacy settings.
Relevant inputs, attached files, and surrounding context can be transmitted to third-party AI providers, each operating under its own privacy notice. Managed services may use Vellum-owned credentials; an own-key connection sends data using the user's credentials under the user's relationship with that provider. Local workspace storage therefore does not, by itself, mean all processing stays on the device.
The security documentation also describes an Ollama option for fully local model inference.
The terms say users retain ownership of their inputs. Subject to compliance with the agreement, Vellum assigns its rights in generated output to the user as between the two parties. The same terms grant Vellum an irrevocable, non-exclusive, non-transferable license to use content to perform the agreement and provide, maintain, and improve the service.
No. Memory stores selected useful knowledge across conversations. The active context is the smaller working set for a particular turn, and long conversations can be summarized when they no longer fit the model's context. Not every sentence becomes a memory or appears in every future response.
No. Desktop can connect to a cloud assistant or a local workspace. A local workspace can still use a hosted model provider, and the computer must stay awake for native local operation. Hosting location and model-inference location are separate choices.
The published packages combine a machine, storage, and a named monthly usage allowance. Usage can cover model calls, search, image generation, and paid service APIs. The reviewed package descriptions do not convert those allowances into a fixed number of completed tasks.
