Vellum
Check which Vellum you are buying: the LLM evaluation platform people recommend is gone, its product pages redirect, and vellum.ai now sells an MIT-licensed personal assistant with eight memory types and a genuinely good permission model. The engineering is serious. The evidence that it works for anyone outside the company is close to nonexistent.
PRICING
| Base | $0 | Free forever. Small machine (1 vCPU, 2 GiB RAM), 6 GiB persistent storage, pay-as-you-go credits with no monthly minimum, managed LLM credentials so you do not need your own API keys |
| Mighty | $30/mo | Entry Pro package: small machine, 10 GiB storage, monthly Mighty usage allowance. The $10/mo platform fee is NOT included, so no custom subdomain, no static IP, no priority support |
| Super | $100/mo | Medium machine (2.5 vCPU, 5 GiB), 30 GiB storage, Super usage allowance, platform fee included: custom subdomain, static IP, priority support, plus assistant email address and subdomain |
| Ultra | $200/mo | Large machine (4 vCPU, 8 GiB), 60 GiB storage, Ultra usage allowance, platform fee and assistant email included |
| Custom | From $50/mo | Pick components: $10/mo platform fee, machine tier $35/$60/$125 (medium/large/XL), storage $5-$30 for 10-120 GiB, optional usage bundle $25/$45/$115. Minimum viable configuration is medium machine plus 10 GiB with no bundle |
| Self-hosted | $0 | MIT-licensed runtime on your own Mac, VPS or private cloud. No platform fee, no machine tier. You supply model API keys and own the uptime |
The pricing is honest and unusually legible for this category: token costs are passed through at provider cost with no markup, the free tier includes managed model credentials so you can start without an API key, and self-hosting removes the vendor from the bill entirely. Two things to model before committing. First, the packages price compute and storage, not usage, so your real monthly cost is the package plus whatever credits an always-on proactive assistant burns; an hourly heartbeat that re-reads notes and background memory filing both consume tokens whether or not you spoke to it that day. Second, there is no seat, no team plan and no admin console anywhere in the pricing. Each subscription buys one assistant tied to one person's accounts, so a five-person team is five subscriptions and five memories that never meet. Prices converted from published USD at 0.86 (ECB, 28/08/2026).
checked 2026-08-31 · vendor pricing page
Element scores
Strengths
Two things here are better than the category norm, and both are structural. The first is the security model. Actor identity is resolved once as guardian, trusted or unknown and enforced everywhere; the sandbox is native OS-level, sandbox-exec on macOS and bubblewrap on Linux, and fails closed if the backend is unavailable rather than quietly running unsandboxed; credentials sit in a separate Credential Execution Service and are injected by a network proxy so the model never sees a raw token; every host action needs permission, granted once, for ten minutes, or always; the default is deny. For software you hand your email, your calendar and your Mac to, that is the right architecture, and it is published rather than asserted. The second is memory. Eight typed memory stores, each with its own staleness window, hybrid dense and sparse retrieval, embeddings on local ONNX by default, per-user and per-channel isolation, and knowledge organised as linked concept articles you can browse and correct rather than an opaque blob. Recall is deliberately bounded to what fits the turn, which is the opposite of stuffing the window and hoping. Around those two, the rest is well built: one assistant reachable from macOS, iOS, web, voice, email, Telegram, Slack and Twilio sharing one memory; an hourly proactive loop that surfaces what is due; a real plugin system on the Agent Plugins 1.0.0 standard with skills, hooks, tools, routes, channels, React apps and MCP servers, distributed through a catalogue that pins immutable commits; broad model choice down to local Ollama, with a Cost profile on a cheap model and an explicit reasoning-off Speed profile. The whole runtime is MIT and actively developed, last pushed the day before this audit, so self-hosting is a genuine exit rather than a marketing line.
Honest dings
Start with the identity problem, because it will cost someone money. Until early 2026 vellum.ai was an LLM development platform for prompt engineering, evaluation and workflow orchestration, and much of the internet still recommends it as one, including a June 2026 review that tests the old product as though it were current. The old product pages now redirect to the assistant homepage while docs.vellum.ai still serves the legacy platform documentation in full, so two different products answer to the same name and the same domain. If you arrived here looking for evals and a prompt playground, you are in the wrong place. Second, the evidence base is thin to the point of being a finding in itself. Independent analysis in August 2026 found no Show HN or Launch HN thread for the assistant, no G2, Capterra or Trustpilot listing, several of the top-ranking comparison threads reading as vendor-seeded, and the widest-reach launch post on X explicitly labelled Sponsored with gift cards offered to early testers. The one substantive Hacker News comment rated it better than the alternative the commenter had tried, then noted a provider bug pushed him to write his own. That is roughly where a product launched in the first half of 2026 should be, but nobody should tell you it has user consensus. Third, no evaluation surface, from the company that used to sell one. Nothing measures whether memory recalled the right thing or a tool did what it claimed, on software with standing permission to send email and run commands. Fourth, no team story at all: no seats, no shared workspace, no admin console, so anything with a procurement process attached stops here. Fifth, the documentation contradicts itself on inference, naming Anthropic with zero retention in the privacy section while the model profiles page lists GPT-5.6 and DeepSeek V4 Flash as the shipped defaults, which is exactly the fact a privacy-conscious buyer needs to be right. Sixth, hosted data lives in Vellum's US infrastructure with no documented EU region, so European buyers who care about residency are pushed to self-host.
Sources (16) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- vellum.ai/docs/pricing — Official pricing documentation, verified 31/08/2026: Base free (small machine 1 vCPU/2 GiB, 6 GiB storage, pay-as-you-go credits, managed LLM credentials), Mighty $30, Super $100, Ultra $200 with machine, storage and usage per package, $10/mo platform fee excluded from Mighty, Custom from $50/mo with itemised machine ($35/$60/$125), storage ($5-$30) and usage bundle ($25/$45/$115) tiers, and the stated at-cost token pass-through policy with no markup (accessed 2026-08-31)
- vellum.ai/pricing — Official pricing page, verified live 31/08/2026: confirms Mighty/Super/Ultra tiers and storage, the assistant email and subdomain entitlement, and the self-hosting statement that the assistant is open source with no platform fee or computer tier charge when deployed on your own hardware, VPS or private cloud (accessed 2026-08-31)
- vellum.ai/docs/trust-security/privacy-and-data — Official privacy documentation: what stays in the workspace (credentials, trust rules, custom skills, config) versus what leaves (model calls, channel messages, tool network calls); Credential Execution Service isolation; Vellum Cloud data held in a dedicated encrypted container in Vellum infrastructure; the self-host opt-out path; and the statement that inference runs on Anthropic with a zero-retention policy, which conflicts with the model profiles page (accessed 2026-08-31)
- vellum.ai/docs/developer-guide/security — Official security documentation and the primary evidence for the Guardrails score: sandbox-exec with SBPL profiles on macOS and bubblewrap on Linux, fail-closed execution, workspace tools confined to ~/.vellum/workspace versus host tools gated by trust rules and permission prompts, and macOS Keychain credential storage with an encrypted file fallback on Linux (accessed 2026-08-31)
- vellum.ai/docs/key-concepts/memory-and-context — Official memory documentation: memory as linked concept articles rather than a transcript, background review of eligible conversations, manual memory creation, selective bounded recall following links between concepts, and the explicit distinction between persistent memory and per-turn context (accessed 2026-08-31)
- vellum.ai/docs/key-concepts/model-profiles — Official model routing documentation: four built-in profiles (Quality on GPT-5.6 Sol, Balanced and Speed on GPT-5.6 Luna with reasoning off for voice, Cost on DeepSeek V4 Flash), Balanced as the workspace default applied to conversations, memory filing and title generation, with per-conversation and per-action-type overrides. Primary evidence for Small models and Thinking models (accessed 2026-08-31)
- vellum.ai/docs/extensibility/mcp — Official MCP documentation: plugins declare servers in a root mcp.json following the Agent Plugins 1.0.0 MCP schema, supporting stdio, SSE and streamable-HTTP transports, connected on install and merged into the same tool catalogue as user-configured MCP servers. Primary evidence for the Protocols score (accessed 2026-08-31)
- vellum.ai/docs/developer-guide/features — Official feature documentation: OAuth2 integrations exposed as bundled skills with a unified messaging layer over Gmail, Slack and Telegram, tokens held in the credential vault and never shown to the model, plus dynamic skill authoring where the assistant drafts TypeScript, tests it in a sandbox, persists and loads it at runtime (accessed 2026-08-31)
- vellum.ai/docs/developer-guide/architecture — Official architecture documentation: Bun and TypeScript runtime owning conversation history and attachments in local SQLite, native Swift macOS client with computer use, and a gateway handling Telegram, Twilio voice, OAuth callbacks and authenticated reverse proxy. Also the source for the SSE event stream and remote access design (accessed 2026-08-31)
- github.com/vellum-ai/vellum-assistant — Official open-source repository, MIT licence, TypeScript, created 07/02/2026, 1,167 stars and 165 forks, last pushed 30/08/2026. README is the source for the eight memory types with staleness windows and per-user isolation, actor identity tiers, hourly proactivity loop, channel list, OAuth provider list, multi-provider model support including Ollama and local ONNX embeddings, the CLI, and the statement that the managed Vellum Platform sustains the business (accessed 2026-08-31)
- vellum.ai/blog/introducing-vellum — Official announcement of the pivot (07/05/2026): the Personal Intelligence positioning, the two-layer memory design of always-loaded working memory plus a selectively retrieved knowledge graph, and the $25M raise from Dharmesh Shah, Arash Ferdowsi, Rebel Fund and Y Combinator (accessed 2026-08-31)
- eesel.ai/blog/vellum-ai-alternatives — Independent analysis (17/08/2026), competitor-authored and treated as adversarial but specific and checkable: confirms the workflow, evaluation, prompt engineering and agent product pages now redirect to the assistant homepage while the developer docs still describe the old platform; documents the absence of any G2, Capterra or Trustpilot listing for the assistant, no Show HN or Launch HN, apparently vendor-seeded Reddit threads, and a Sponsored launch post on X offering gift cards to early testers (accessed 2026-08-31)
- eesel.ai/blog/vellum-ai-pricing — Independent pricing and evidence analysis (18/08/2026): the same competitor source on the name collision between three products, the thin community footprint, the single substantive Hacker News comment about the assistant, and the structural finding that Vellum prices one assistant package with no seats, shared workspace or admin console, so a five-person team means five subscriptions and five separate memories (accessed 2026-08-31)
- cybernews.com/ai-tools/vellum-ai-review — Independent review (17/06/2026) that tests the legacy LLMOps platform as if current, and is itself evidence of the confusion: flags the five-user cap on paid plans, the engineering setup requirement, and explicitly names overlapping branding between the AI assistant and the LLMOps platform as making pricing and product details hard to find (accessed 2026-08-31)
- docs.vellum.ai/llms.txt — The legacy platform documentation still served in full on the docs subdomain as of 31/08/2026, covering prompt engineering, workflows, nodes, evaluation and deployments. Primary evidence that both products remain live under one brand (accessed 2026-08-31)
- news.ycombinator.com/item?id=35042836 — Launch HN, 06/03/2023: founders Noa Flaherty, Akash Sharma and Sidd Seethepalli introducing Vellum (YC W23) as a developer platform for LLM apps with playground, semantic search, monitoring and fine-tuning. Establishes the original category the brand still ranks for (accessed 2026-08-31)