Skip to main content

Why Vanta

Vanta is a local, trusted-operator agent: it knows the goal before it picks a tool, enforces scope on every action, and reports only verified output. It runs on your hardware, on any model, and the safety boundary travels with it.

Here's what that buys you — each point is a concrete, checkable claim, not a slogan.

A boundary the agent can't talk its way past​

Vanta's Rust kernel risk-classifies every action into allow / ask / block, enforces scope containment, and writes a tamper-evident audit chain — and the agent loop cannot bypass it. Execution runs inside an OS sandbox, Docker, or SSH backend. The safety boundary is a separate enforced layer, not a feature the model can reason around.

It knows the goal before it acts​

A goal ledger plus a dependency graph mean Vanta carries what it's trying to do into every step — so an action that doesn't serve the goal is visible as such, before it runs. See the agent loop.

Built around executive function, not against it​

Most agents assume the bottleneck is capability. For a lot of people — especially neurodivergent operators — the real bottleneck is coherence: starting, not drifting, and closing a loop before opening the next one. Vanta is ND-first by design, with executive-function support as a first-class layer, not an afterthought:

  • Automatic adaptation — observable task cues such as stuckness, corrections, overload, and reorientation needs privately reshape the current turn without inferring a diagnosis or writing a permanent profile fact.
  • Self-redirect — repeated failures and research-only action loops receive one bounded change-of-approach directive before the existing hard stop.
  • Task initiation — stuck or staring? Vanta prescribes the single smallest concrete next step, not vague encouragement.
  • Choice reduction — a backlog surfaces as the top 3 by effort, smallest first, with the rest hidden — never a wall of options.
  • Working memory — it re-anchors the goal and current step across a long session, and after any context compression, so you don't lose the thread.
  • Closure over novelty — gates flag unfinished, uncommitted work before you context-switch, and bias toward "where do we stand?"
  • Time-blindness help — estimates come as best / realistic / worst ranges with named hidden costs, never a single misleading number.
  • Low sensory load — compressed output by default: one idea per line, the single most important thing in bold.

See executive function.

It learns you, on your machine​

vanta tune lora trains a local adapter from the operator decisions you've actually accepted and rejected — real on-device (MPS / CUDA / CPU) training, nothing leaves the machine. Plus auto skill-improvement, reflection-from-correction, and an evolving memory. See self-improvement.

Reachable from anywhere you already are​

20 messaging adapters from one gateway — Telegram, Slack, Discord, Signal, WhatsApp, iMessage, Matrix, LINE, Mattermost, IRC, ntfy, Teams, Twitch, SMS, Zalo, Feishu, WebChat, Nostr, Google Chat, Email (5 verified live end-to-end; the rest need only your credentials). See comms & gateway.

Any model, any host you control​

Provider-agnostic: any OpenAI-compatible endpoint (custom base URL) plus Azure, OpenRouter (200+ models), Ollama, OpenAI, Anthropic, and Gemini — and Vanta exposes its own OpenAI-compatible proxy. Run it local, in a sandbox, in Docker, over SSH, or on a $5 VPS; the kernel enforces scope identically everywhere. See providers.

Yours, not rented​

MIT-licensed and fully self-hosted. Your data residency, no vendor lock-in, no hosted service in the loop.

Honest about the bar​

Some capabilities are live-proven today; others are built and need one-time setup for live use (a provider key, OAuth client, browser binaries, or a platform credential). We label which is which rather than implying everything is turnkey — see what works now and each capability page. The claims above are things you can verify by running it, not marketing superlatives.

Ready? Quickstart gets you talking to Vanta in one command.