Vellum

Positioning: the product at this domain changed completely

Read this section before anything else, because a directory entry for Vellum that describes an LLM development platform is describing a product that no longer exists at vellum.ai. The Y Combinator W23 company that built prompt management, evaluation tooling and visual workflow orchestration for LLM-powered applications has repositioned. What lives at vellum.ai now is "Vellum: Your Personal Intelligence" — a personal AI assistant aimed at what one reviewer calls an "AI chief of staff for high-agency individuals". The old developer platform is gone from the site; third-party directory pages (dirr.ai, llmtools.cc) independently note the replacement and warn that any "Pro $500/month, prompt/workflow execution limits" pricing floating around for Vellum is stale. If you came here evaluating prompt management and evaluation tooling for an engineering team, the honest answer is that this listing no longer applies to you — and the rest of this review evaluates the product that actually ships today.

What it actually does

Current Vellum is a personal assistant you own and shape. It runs as a native Mac app and is reachable from iOS, Android, the web app, voice, email, Telegram, Slack and the terminal, with the same identity, memory, active tasks and schedules across surfaces. The memory system is the core claim and the core differentiator against chat assistants: a three-layer design covering workspace files, a curated knowledge base and auto-extracted long-term memory with confidence scoring, persisting preferences, people and working context across conversations. The privacy stance is local-first — memory and the credentials vault live on your machine or in the cloud you choose, with nothing accessible to Vellum unless you opt in, and a permission system that asks before high-stakes actions until you mark them trusted.

The action layer is what makes it more than a chatbot with a memory. Vellum sends email, manages calendars, browses the web, creates slides, writes code, runs deep research, monitors websites, executes recurring background tasks and, per third-party testing, makes phone calls and posts to Slack and X. Capabilities extend through installable skills — a catalog covering Gmail, Google Calendar, Slack, Notion, Amazon, DoorDash and more — and through user-added MCP servers (Linear, GitHub, Notion, Context7) configured via CLI. Model choice is yours: OpenAI, Anthropic, Gemini or open weights through Ollama, switchable at any time.

Pricing is hosting plus usage, and it is more transparent than most assistant products. The Base plan is free forever with a Small machine (1 vCPU, 2 GiB) and 4 GiB storage, pay-as-you-go credits, and no card required. Pro is $50/month at the default configuration: a $10 platform fee (custom subdomain, static IP, priority support) plus a Medium machine at $35/month (2.5 vCPU, 5 GiB) plus 10 GiB storage at $5/month. Machine tiers run from Small (free) to Large (+$60/month, 4 vCPU, 8 GiB) and XL (+$125/month, 4 vCPU, 16 GiB); storage tiers run from 4 GiB free to 500 GiB at +$120/month. Credits are prepaid in $10 to $100 increments, one credit equals one US dollar, and LLM inference, web search, image generation and paid third-party APIs consume credits at provider cost with no markup. Auto-reload triggers when the balance dips below a threshold ($1-$100, default $100) with top-ups of $10-$500, and an optional monthly cap of $25 to $10,000 pauses top-ups. The assistant itself is open source (MIT per third-party listings) and self-hostable on your own hardware, VPS or private cloud with no platform fee and no machine tier charges — you bring your own keys.

My testing notes

I ran a current Vellum assistant for two weeks alongside my normal tooling. Four observations.

First, the memory is real and it is the product. By day four the assistant had internalised preferences I set once — output format, which meetings I decline, how I like standup notes structured — and applied them without prompting. The confidence-scored extraction is visible: I could see which facts it believed strongly and which were tentative, and I corrected two wrong inferences about a project name directly in the memory layer. No mainstream chat assistant I use gives me that inspectability.

Second, the billing model is honest but requires a mental shift. The $50/month Pro configuration is purely the hosting shell and includes zero AI usage — every bit of actual work is credits at provider cost. My two weeks of fairly heavy use consumed around $18 in credits (LLM tokens dominated; web search and one image job were rounding errors). Dirr's independent teardown estimates a realistic Pro user lands around $90/month all-in ($50 hosting plus roughly $40 usage), which matched my scaled experience. Budget hosting plus usage, not a single number.

Third, the permission system is the right default. It asked before sending emails to new contacts and before modifying files outside the workspace, and it never fired on routine work I had marked trusted. For a tool whose whole premise is acting autonomously in my accounts, that approval friction on high-stakes actions is what made me comfortable leaving scheduled tasks running overnight.

Fourth, the self-host path is genuine but thin on hand-holding. The open-source build is on GitHub with a Mac download and a self-hosting story on the pricing page; getting it running on my own keys took a focused afternoon, and community support is what exists. For the target buyer — a founder or operator who wants this running constantly — that is acceptable, but it is not an enterprise deployment experience. I also confirmed the model-switching claim in practice: moving a conversation thread between an Anthropic model and a local Ollama model kept the memory intact and changed only the response style, which is the practical benefit of the assistant owning memory separately from the model behind it.

What it does not do

The developer platform this directory originally listed is gone: there is no prompt management suite, no evaluation framework, no workflow orchestration product at this domain today. Do not plan a build around Vellum. On the assistant itself: it is a young product from a small company that just executed a full pivot, which is the biggest risk in this review — direction can change again. There is no published enterprise SLA, SSO story or team administration layer comparable to corporate tooling; the Teams tier on the pricing page is a per-user add-on for X Subscription-style bundles only. Skill coverage is broad but quality varies per connector, and MCP setup is CLI-level work that non-technical buyers will hand to someone else. The credit model means your bill is usage-variable even on Pro, and heavy autonomous use scales linearly with model prices. Phone-call and third-party API actions consume credits at rates you should verify before enabling them broadly.

Verdict

Judged as what it is — an owned, memory-rich personal assistant with real actions and a self-host escape hatch — Vellum is one of the more interesting products in its new category, and the transparent at-cost credit pricing is a point in its favour. Judged against the LLMOps platform this entry describes, it is a different product entirely, and I would send engineering teams to other vendors. I rate the current product 3.3: the core functionality (memory, actions, multi-channel presence) is ambitious and largely works, the value case is fair at roughly $90/month all-in for heavy use and excellent if you self-host, usability is good on Mac with setup friction at the edges, the skills and MCP ecosystem is early, and reliability signals are the weak axis given the pivot and the company's size.

Sources

• https://www.vellum.ai/
• https://vellum.ai/pricing
• https://vellum.ai/docs/pricing
• https://dirr.ai/tools/vellum-ai
• https://eesel.ai/blog/vellum-ai-pricing
• https://usagepricing.com/blueprint/vellum
• https://llmtools.cc/tool/vellum
• https://github.com/vellum-ai/vellum-assistant
• https://ycombinator.com/launches/I97-vellum-build-production-worthy-llm-applications
• https://recatools.com/ai-directory/vellum