Vellum
Also known as: Vellum Assistant, Vellum Cloud, vellum.ai
Vellum is an open source personal AI assistant that does one person's busywork: it sends email, manages calendars, browses the web, orders and books, makes phone calls and runs recurring jobs in the background. Each assistant keeps long term memory of its owner's people, preferences and work, asks before higher stakes actions such as emailing a new contact or spending money, and runs on a Mac or Windows PC, in Vellum Cloud from $30 a month, or on the owner's own server.
Vellum, the New York company led by chief executive Akash Sharma, now sells one product: a personal AI assistant for the busywork of one person's life and job. The assistant has its own identity, its own email address on the larger cloud plans and a workspace of files, notes and skills, and it takes actions instead of answering in a chat window. The code is open source under the MIT license, and the same assistant runs on a Mac or Windows PC, in Vellum Cloud, or on a server the owner runs.
Out of the box the assistant connects to Google (Gmail, Calendar, Drive, Docs, Sheets and Contacts), Outlook, Slack, Notion, GitHub, Linear, Asana, HubSpot, Todoist, Discord and X, and a library of 76 skills adds Amazon shopping, DoorDash food and grocery orders, Twilio phone calls, image generation, transcription and a morning briefing. A headless browser fills forms and signs in with stored credentials, and computer use drives desktop apps through screenshots and the accessibility tree.
Work can start without a prompt, from schedules written in plain language, an hourly heartbeat checklist, watchers that poll Gmail, Outlook, GitHub, calendars and Linear, and playbooks that act on incoming messages. Bigger jobs go to subagents that run in parallel as researchers, builders or advisors. Memory is a linked wiki of concepts about the owner's people, preferences and work, browsable as a map, corrected with Refine and deleted on request. The owner picks the model through inference profiles, per conversation or per action, across OpenAI, Anthropic, Google, OpenRouter, Fireworks and local models through Ollama.
Downloading and self hosting are free, with no platform or computer fee. Vellum Cloud runs each assistant in its own encrypted container on Mighty at $30 a month, Super at $100 or Ultra at $200, each with a dedicated computer, storage and included usage, and credits at $1 each pay for model tokens at cost, web search, image generation and paid outside APIs.
Every action carries a risk tier, and the owner sets how much runs without asking, separately for chat, scheduled work and API calls. Anything else stops at an Allow or Deny prompt that can become a saved trust rule. Purchases go through Stripe's Link wallet, each approved checkout capped at $500. Credentials sit in an isolated vault and are injected into requests, so the model never sees them. There is no security certification, and the cloud plans offer no choice of region.
Vellum fits a professional who wants one assistant across email, calendar, messaging and errands, with open code and the option to keep everything on a personal machine or home server. It suits people comfortable choosing models and approving actions, and developers who want to extend it with their own skills, plugins and MCP servers. It fits less well for a company rolling assistants out to staff, with no team accounts, single sign on or compliance report, and for anyone who wants to test how the assistant behaves before a change goes live.
Vendor details
Canonical URL
https://www.vellum.ai
Category
Enterprise operations agent
Subcategory
Open source personal AI assistant that handles email, calendar, errands and recurring background work, locally, in Vellum Cloud or self hosted
Funding status
Independent, led by chief executive Akash Sharma with co chief technology officers Noa Flaherty and Sidd Seethepalli.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Google (Gmail, Calendar, Drive, Docs, Sheets, Slides and Contacts in one connection), Outlook and Microsoft calendar, Slack, Notion, GitHub, Linear, Asana, HubSpot, Todoist, Discord and X connect through OAuth, with tokens kept in the workspace's own vault. Skills add Amazon, DoorDash, Twilio phone calls, AgentMail email, image generation and Whisper transcription, and web search runs through Vellum, the chosen model's own search, or the owner's keys for Perplexity, Brave Search, Tavily, Firecrawl, Exa or Serply. Purchases go through Stripe's Link. Plugins add tools, hooks, routes, channels and MCP servers, and other services connect through a custom skill with an API key.
In practice
A founder asks Vellum to watch the inbox overnight. A Gmail watcher picks up each new message, the assistant drafts replies for known contacts, holds anything addressed to someone new until the founder taps Allow in Slack, and the morning briefing lists what is waiting.
A parent asks the assistant to restock the pantry for the weekend. It builds a DoorDash grocery order and sends a Link spend request naming the merchant, the amount and the purpose, and nothing is charged until the request is approved in the Link app.
A developer runs the open source assistant on a Mac Mini at home, sets a budget profile for conversation titles and a stronger model for the main agent, and adds an MCP server and a custom skill for a home automation API, so only model calls leave the house.
Sources & related URLs
Related / legacy domains
Research sources
Agentic Index coverage score
11.5 / 14 capabilities · 82%
| Integrations & Tool Calling | Full |
|---|---|
|
OAuth connections let the assistant act in Gmail, Google Calendar, Drive, Docs, Sheets, Slides and Contacts through one Google sign in, and in Outlook, Slack, Notion, GitHub, Linear, Asana, HubSpot, Todoist, Discord and X, where it sends messages, opens issues, updates tasks and deals, and posts. Scopes are kept to what each service needs, and tokens stay in the workspace's own vault, injected into each request so the model never sees them. A service outside the list connects through a custom skill with its API key, and plugins can register MCP servers whose tools join the catalog. Sourcevellum.ai/docs/key-concepts/oauth-integrationsread 2026-11-19 |
|
| Workflow Orchestration | Full |
|
Larger jobs split across subagents that run in the background, several at once, while the main conversation carries on. A regular subagent starts fresh with only its objective, a fork inherits the whole conversation and its memory, and an advisor works from a brief on a stronger model without changing anything. Roles set what each may do: researchers only read, builders write files, run commands and use connected apps, and an unrecognized role falls back to read only. Each subagent moves through pending, running and completed, failed or aborted, can raise interim findings or say it is blocked, and the parent is told when it finishes. Schedules can also start a saved workflow. Sourcevellum.ai/docs/skills-reference/subagentread 2026-11-19 |
|
| Knowledge Grounding & RAG | Partial |
|
Answers draw on context gathered for each task: Google Drive, Docs and Sheets, Notion pages, files in the assistant's own workspace, and live web search with cited sources. Search runs on Vellum's managed service, on the chosen model's own search, or on the owner's keys for Perplexity, Brave Search, Tavily, Firecrawl, Exa or Serply, with a fallback chain across connected keys, and Firecrawl also reads full pages that render with JavaScript. No maintained index over the owner's documents persists between runs to be searched as a library; the long term store is the memory wiki of facts about the owner. Sourcevellum.ai/docs/key-concepts/web-searchread 2026-11-19 |
|
| Human Oversight & Guardrails | Full |
|
Every action carries a low, medium or high risk tier, and the owner's risk tolerance (Strict, Default, Relaxed or Full access) decides what runs without asking, with separate thresholds for chat, scheduled tasks and calls through the API. Anything above the line stops at an Allow or Deny prompt showing the action, its risk badge, the reason and the full tool input, and a denied action is not retried without consent. Approvals can become trust rules by tool, pattern and scope, with deny winning a tie. Purchases wait for the owner's approval in the Link app before any charge, and approvals sent to Telegram or Slack expire after five minutes. Sourcevellum.ai/docs/trust-security/the-permissions-modelread 2026-11-19 |
|
| Security, Identity & Governance | Partial |
|
Contacts are the access list for anyone other than the owner. A contact reaches the assistant only through a verified channel on Telegram, Slack or phone, invite links are channel specific and revocable, each channel can be active, revoked or blocked, and a trusted contact can chat but needs the owner's approval before it touches memories or sensitive tools. Credentials sit in a separate Credential Execution Service, encrypted at rest and injected into requests so neither the assistant nor the model sees raw values. There are no team accounts, roles or single sign on, and the assistant comes with no SOC 2 report or other security certification. Sourcevellum.ai/docs/skills-reference/contactsread 2026-11-19 |
|
| Observability & Auditability | Partial |
|
Each conversation shows the assistant's tool calls, their results and its approval prompts, and the event stream carries text, tool use, tool results and confirmation requests to any client the owner connects. Subagents report their status and interim findings, the vellum logs and doctor commands show the runtime's own logs and health, and every Link purchase appears in the Link history labeled with the assistant that made it. The owner reviews what happened conversation by conversation, with no single activity log or audit trail that pulls every action across chats, schedules and channels into one record. Sourcevellum.ai/docs/developer-guide/apiread 2026-11-19 |
|
| Memory & State Persistence | Full |
|
Memory is a linked wiki of concept articles about the owner's people, preferences and work. The assistant saves a fact when told something that will matter later, a background review of eligible conversations adds durable facts missed during the turn, and the owner can add one with Create memory. Facts carry across conversations long after the original messages leave the context window. The memory view lays out concepts and their links on a searchable map, Refine corrects a concept so the new version replaces the old, any single memory is deleted on request, and switching memory off stops new capture. Self hosted assistants keep the store on the owner's own machine. Sourcevellum.ai/docs/key-concepts/memory-and-contextread 2026-11-19 |
|
| Deployment & Data Residency | Full |
|
The same assistant runs in three places. The native Mac or Windows app keeps the workspace on the owner's computer with direct access to local files, available while the machine is awake. Vellum Cloud runs each assistant around the clock in its own sandboxed, encrypted container on Vellum's infrastructure, with no choice of region. Self hosting puts it on infrastructure the owner controls, with a step by step guide for a Compute Engine virtual machine in the owner's own Google Cloud project and a Mac Mini at home as the other example, reached from other devices through pairing and tunnels. Prompts still go to whichever model provider is configured. Sourcevellum.ai/docs/hosting-optionsread 2026-11-19 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
A library of 76 skills across email, calendar, messaging, browsing, productivity, development, voice, commerce, content, health and system work installs with a single command, and each skill does its own job: Gmail, Slack, Google Calendar, Notion, Outlook, Amazon shopping, DoorDash orders, Twilio phone calls, image generation, transcription, a morning briefing and a tracker for messages awaiting replies. A curated plugin catalog, reviewed entry by entry and pinned to a fixed commit, adds packaged tools, hooks and MCP servers, and a procedure the assistant completes can be saved as a skill the owner reviews before keeping it. Sourcevellum.ai/skillsread 2026-11-19 |
|
| Triggers & Channel Coverage | Full |
|
Work reaches the assistant without a prompt in several ways. One time schedules fire at a set timestamp, and recurring ones follow cron or calendar recurrence rules described in plain language, in modes that wake the assistant, send a plain alert, run a shell script or start a saved workflow. A heartbeat works through a checklist every hour by default within set active hours, watchers poll Gmail, Outlook, GitHub, Google and Outlook calendars and Linear as often as every 15 seconds without handling an event twice, and playbooks act on incoming messages that match a trigger. Messages also arrive by Telegram, Slack, email, phone and plugin webhooks. Sourcevellum.ai/docs/key-concepts/schedulingread 2026-11-19 |
|
| Model Flexibility & Routing | Full |
|
The owner chooses the model. Inference profiles set a workspace default, with built in Quality on GPT 5.6 Sol, Balanced and Fast on GPT 5.6 Luna and Budget on DeepSeek V4 Flash, and custom profiles add any other provider or model. A picker changes the profile for one conversation, and action overrides give each job its own profile, such as the main agent, memory extraction or conversation titles. Vellum supplies credentials for its managed models, the owner's own keys open OpenAI, Google Gemini, OpenRouter and Fireworks, and local models run through Ollama. Sourcevellum.ai/docs/key-concepts/model-profilesread 2026-11-19 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
The runtime serves a versioned REST API behind JWT bearer tokens signed by its gateway, which lists conversations, posts messages and streams model output, tool calls, approval prompts and channel events over server sent events. The vellum command line tool sends messages, tails events, reads logs and points at a remote assistant, and plugins extend the assistant with new tools, lifecycle hooks, HTTP routes under the plugin's own path, channels, apps and MCP servers. The assistant, gateway, command line tool and credential service are open source under the MIT license, so a developer can build on or change any part. Sourcevellum.ai/docs/developer-guide/get-startedread 2026-11-19 |
|
| Testing, Debugging & Optimization | Not documented |
|
The assistant comes with no way to test its own behavior. There is no test suite, scored run, simulation or side by side comparison of models or profiles that an owner can point at the assistant's work. Choosing a profile rests on guidance, which recommends stronger models for anything shown in a conversation or driving a decision and lighter ones for short structural jobs such as titles and starters. Self improving skills turn a completed procedure into a skill the owner reads and approves before it joins the workspace. Sourcevellum.ai/docs/key-concepts/model-profilesread 2026-11-19 |
|
| Browser & Computer Use | Full |
|
Computer use drives desktop apps the way a person does. On each step the assistant reads the accessibility tree, clicks controls by name or by screenshot position, types and scrolls, then takes a screenshot to see what the action did. It works in the Mac and Windows apps, and conversations on the hosted web app get a streamed virtual desktop. Each action asks for approval unless a trust rule allows it, and a session stops at 50 actions. A separate headless browser in the sandbox navigates pages, fills forms, takes screenshots and signs in with credentials stored for that domain. Sourcevellum.ai/docs/skills-reference/computer-useread 2026-11-19 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Version 0.12.8 of Vellum lets users export a platform hosted assistant directly from the backups card in the web interface. The local runtime can now stream import platform bundles during teleport, so an assistant can be migrated from Vellum's cloud to a locally hosted runtime end to end without manual steps, with credentials stripped from the bundle on import.
Bears on: Deployment / data residency
View sourceVellum moved to wide release with plugins as a first-class marketplace, a new Advisor capability that pulls in a second more powerful model on hard problems, Memory v3 at section grain for sharper recall, and upgrades to subagents, workflows, Slack, and the Activity page.
Bears on: Agent capability
View sourcePricing
Free to start; Mighty $30 a month; Super $100; Ultra $200; custom plan
A monthly plan for the cloud computer and storage, plus prepaid credits for usage beyond what the plan includes
What is public
Vellum is free to download and run on the owner's own computer or server, with no platform fee and no computer charge. Vellum Cloud sells three monthly plans, each with a dedicated computer, storage and included usage. Mighty is $30 a month with 1 vCPU, 2 GiB of memory and 10 GB of storage. Super is $100 a month with 2.5 vCPU, 5 GiB and 30 GB, and Ultra is $200 a month with 4 vCPU, 8 GiB and 60 GB. Super and Ultra also give the assistant its own email address and subdomain. A custom plan is built to order.
Billing mechanics
One credit costs $1. Credits pay for model tokens, passed through at cost, for searches on Vellum's own search service, for image generation and for paid outside APIs such as X, and they are drawn only after a plan's included usage runs out. Credits are bought in top ups of $10 to $100 through Stripe Checkout, or reloaded automatically when the balance falls below a set threshold, with an optional monthly cap from $25 to $10,000.
Cost watchouts
Model tokens make up most usage, so steady work on stronger models raises the credit bill. Super and Ultra carry a $10 monthly base fee for a custom subdomain, a static IP address and priority support. Searches through the owner's own Perplexity, Brave Search, Tavily, Firecrawl, Exa or Serply keys are billed by those providers directly.
Sales call required
Mixed (some tiers require a call)
Free / trial
A free start on Vellum Cloud, with the first tasks included
Cancellation / refund
Plans can be cancelled or changed at any time. Upgrades take effect at once, computer and storage changes are prorated, and downgrades or cancellations take effect at the end of the billing period. Purchased credits expire 12 months after purchase and are not refundable except where the law requires it.
Missing data
Vellum does not say how much usage each plan includes, and it offers no annual price.
Related vendors
- 11th Estate — Agentic platform whose AI engine scans markets, matches an…
- Abacum — FP&A platform for growing companies whose AI Modeler builds, edits…
- Accend — Accend builds human in the loop AI agents for commercial credit…
- Adopt AI — Finance agents that categorize, reconcile and close the books inside…
- Aera Technology — Decision intelligence platform where an always on agent executes…
- Akro AI — On premise operational intelligence that automates document heavy…
Alternatives to Vellum
Other products that do the work Vellum does, and what each one is.
- PallyFree assistant you text
- PokeProactive assistant in messages
- Meta MuseMeta's agent in a cloud VM
- Tasks in Gemini (formerly Gemini Spark)Background tasks in Gemini
- ChatGPT dotsAlways on agent in ChatGPT
- PerplexityCloud digital worker