Vellum AI

Vellum is now an open-source personal AI assistant with memory, skills, integrations and optional computer access. It is designed to become a persistent digital colleague that can handle email, coordinate work across tools, maintain repositories, update project systems and run scheduled tasks. This is a major product change: older descriptions of Vellum as an enterprise LLM application-development platform no longer reflect the current public product.

What is Vellum today?

The current Vellum Assistant has its own workspace, identity, long-term context and permission model. Users can connect tools, install plugins and skills, choose model profiles, create schedules and interact through supported channels. The assistant is intended to learn preferences over time and take action, not merely answer isolated prompts.

Vellum is available as a managed cloud service and as open-source software that can run on a user’s Mac, server or private cloud. Self-hosting removes Vellum’s platform and computer-tier charges, but the user becomes responsible for infrastructure, updates, backups, security and provider costs.

Current Vellum pricing

• Free: a starter allowance for trying the assistant and its first tasks.
• Mighty: $30 per month with 1 vCPU, 2 GiB memory, 10GB storage and included Mighty usage.
• Super: $100 per month with 2.5 vCPU, 5 GiB memory, 30GB storage, included Super usage, an assistant email address and subdomain.
• Ultra: $200 per month with 4 vCPU, 8 GiB memory, 60GB storage, included Ultra usage, email and subdomain.
• Custom: configurable capacity outside the standard packages.

Plans can be changed or cancelled. Managed packages combine hosted-computer capacity, storage and usage. Heavy browser work, coding, background schedules and larger models can require more resources than conversation.

Model costs and credits

Vellum states that it passes through model-provider costs without adding a margin to token usage. Credits pay for models and AI services called by the assistant. Users can buy credits and configure spending controls or automatic reloads.

The monthly package pays for the managed environment and its allowance, while consumption depends on the model, context length, voice, image, search and automation activity. Self-hosting removes the Vellum platform fee but does not make third-party services free.

Core capabilities

Persistent memory and context

Vellum accumulates context about the user, role, preferences and ongoing work. Memory can reduce repetitive setup and support proactive follow-up. It also increases privacy sensitivity, so users should review what is remembered and remove obsolete or inappropriate material.

Skills, tools and plugins

Skills encode repeatable behaviour, while tools let the assistant interact with files and services. Vellum maintains plugins across productivity, developer, lifestyle and marketing categories and supports extensibility through hooks, MCP, routes, channels and apps.

Every integration expands risk. Install only necessary plugins, inspect permissions and prefer narrowly scoped credentials. A published plugin should not automatically be trusted with confidential data.

Computer use and local files

The assistant can interact with a computer and local files when authorised. Permission choices range from asking before actions to broader autonomy. Start with the strictest profile. Require confirmation for messages, deletion, access changes, purchases, deployment and external publishing.

Email, Slack, GitHub and Linear workflows

Vellum presents workflows for triaging inboxes, drafting replies, summarising Slack, maintaining GitHub issues and updating Linear. Each automation should define what may be drafted, changed or sent without approval.

Schedules and channels

Scheduled tasks allow background work, while channels expose the assistant through other interfaces. Useful examples include morning briefs and project monitoring. Scheduled autonomy needs ownership, audit logs, spending limits and a clear disable procedure.

Model profiles and bring-your-own keys

Model profiles let users choose providers for different work. Vellum’s privacy notice lists providers spanning language, voice, image and search services, and users can configure their own credentials. Bring-your-own-key setups make the user’s contract with that provider directly relevant.

Self-hosting

Vellum Assistant is open source and can be deployed locally, on a VPS or in a private cloud. Self-hosting offers control over infrastructure, data location, keys and uptime. Credentials can be stored in macOS Keychain in supported local setups.

Control is not automatic security. Operators must patch dependencies, protect the network, manage secrets, back up data, monitor logs and secure integrations. Exposing an assistant publicly without hardening may create more risk than the managed service.

Privacy and data use

Vellum says conversations reach model providers for responses but are not used for training AI models. Telemetry is described as off by default. Its privacy notice says that, depending on user preferences, Vellum can collect conversation content and files to evaluate and improve the service without training models; this can be changed in Permissions & Privacy.

Relevant inputs, attachments and context are transmitted to the third-party provider performing a task. Each provider has its own privacy policy. Managed Vellum may route requests with Vellum-owned credentials, while bring-your-own-provider connections fall under the user’s direct agreement.

Review both Vellum and provider policies. Sensitive organisations need answers on retention, deletion, subprocessors, regional transfers and incident handling.

Secrets and permissions

Vellum states that passwords and credentials are stored in macOS Keychain when self-hosted or an isolated vault on the managed platform, with a deterministic service executing them so the AI does not directly see or store secrets. Users must still protect sessions, recovery methods and plugin scopes.

The permission model includes Strict, Conservative, Relaxed and Full Access modes. Full Access belongs only in isolated, well-understood environments. Most users should require approval for communication, purchases, account changes, code deployment and destructive operations.

Who Vellum is best for

• Technical users wanting an open-source, extensible assistant.
• Professionals coordinating email, Slack, GitHub and Linear.
• Users who value memory and scheduled background tasks.
• Teams capable of operating a secure self-hosted deployment.
• People who want model choice and their own provider keys.

Limitations

• The product changed substantially, making older reviews obsolete.
• Managed plans cost more than basic chat subscriptions.
• Provider usage creates variable costs.
• Persistent memory and integrations increase privacy exposure.
• Computer use and schedules require monitoring.
• Self-hosting transfers security responsibility to the operator.

How to evaluate Vellum safely

1. Start with non-sensitive data and strict permissions.
2. Connect one low-risk service before email, files or source code.
3. Test a repeatable task for a week and measure corrections and spend.
4. Review memory, logs, jobs and credentials regularly.
5. Compare managed cost with self-hosting maintenance time.
6. Create approval gates for messages, deployments and destructive actions.

Best alternatives

ChatGPT and Claude are simpler for general conversation. OpenClaw is a direct comparison for open personal agents. Open WebUI suits self-hosted model interfaces. Microsoft Copilot and Google Gemini integrate with their productivity suites. n8n provides more deterministic business automation.

Verdict

Current Vellum is an ambitious personal AI system, not the LLM development platform many older pages describe. Its appeal lies in persistent context, plugins, computer use, schedules, model choice and open-source self-hosting. Those capabilities make permissions and cost management central. Start with a narrow workflow, keep autonomy low, inspect every provider and move to paid or self-hosted deployment only after it proves reliable.