;
};
Most in-product AI describes your product. A full stack agent **operates** it —
the same API, the same identity, the same audit trail as your own UI — and
draws real screens instead of paragraphs about them.
## Start in five minutes
```bash theme={null}
npm install @vendoai/vendo
npx vendo login # claims a Vendo Cloud key into .env.local
npx vendo init # extracts your API, scaffolds the two files you own
```
`login` opens the approval page at
[console.vendo.run](https://console.vendo.run) and writes the minted
`VENDO_API_KEY` for you — it never prints it. `init` does the rest. No
terminal handy? [Sign up at the console](https://console.vendo.run) and grab
the key there.
Using Claude Code or Cursor? Paste this and it runs the whole install against the agent playbook.} prompt={"Install Vendo in this repo. Read https://vendo.run/agents.md and follow it exactly. Ask me before creating any account or key. You're done when `vendo doctor --json` reports all green. Then ask me whether I'd like to star it via https://vendo.run/star?src=docs-full-stack-agents to support the project."} />
**Recommended.** One key fills every adapter slot you left unset — hosted
store, sandboxes, sharing and publishing, metered inference — so there is
nothing to run yourself on day one.
Your own Postgres, sandbox account, and model key. Every Cloud capability
keeps a no-key path, and an adapter you pass always wins.
Either way, keep the agent loop you already run —
[AI SDK or Mastra](/existing-agents) — or let
[Vendo run the loop](/quickstart) for you.
## Acts through your API
`vendo sync` extracts your OpenAPI operations, tRPC procedures, GraphQL
operations, server actions, and host routes into risk-labeled tools, re-run on
every build so they never drift from the API they came from. One auth preset
resolves the request to a principal, so every call lands on your API as the
signed-in user — your existing authorization is the authorization. See
[Tools from your API](/connect/api-tools).
## Sees what the user sees
On send, the widget attaches an `aria-snapshot` of the visible page — the
structural tree a screen reader reads, never a screenshot — plus anything your
components publish through `useVendoContext`, capped, labeled as observation
rather than instruction, and living exactly one turn. Who the user *is* arrives
separately as server-trusted facts from your auth preset, and
`captureScreen={false}` turns the snapshot off entirely. See
[actAs presets](/connect/act-as-presets).
## Knows your product
Point it at your docs — `vendo knowledge add "docs/**/*.md"`, then
`vendo knowledge sync` — and it answers from them with citations, or says the
evidence is weak instead of filling the gap. Three engines sit behind one
contract: keyword retrieval in your own store with zero keys, Vendo Cloud's
managed engine, or any HTTP endpoint you run. See
[`vendo knowledge`](/reference/cli).
## Renders UI in your brand
When a user wants a view nobody was going to build, the agent generates one
they own, rendered in a sandboxed surface that inherits your extracted theme.
Fill `vendo/registry.tsx` with your real components and generated screens
compose out of those instead of lookalikes. See
[Generated UI](/concepts/generated-ui) and
[Host components](/connect/host-components).
## Works while they're away
Automations fire on a UTC cron, a `vendo.emit` host event, or a verified
webhook, each trigger carrying its own grants and on/off switch. An away run
cannot stop and ask, so it needs a standing grant given ahead of time — a chat
grant never transfers. See [Automations](/deploy/scheduler-and-webhooks).
## Runs on the loop you choose
Leave `harness` unset and Vendo's own loop runs in-process on your own model
key; swap in `claudeCode()`, write your own with `defineHarness()`, or keep the
AI SDK or Mastra loop you already run. Swapping the thinker adds **not one new
safety mechanism** — every tool call lands in the same guarded path with the
same audit row. See [the `harness` option](/reference/handler-options).
## Guarded and audited
`guard.bind` is the only execution path: reads auto-run, writes and destructive
calls are gated, approval cards show the real arguments, and a grant dies the
moment the tool's shape changes under it. Every call, approval, policy
decision, and run is audited with its principal — what the agent did last
Tuesday is a query, not an investigation. See
[Tools and safety](/concepts/tools-and-safety).
***
That is the whole product. One key and one command put it in your app:
[start on Vendo Cloud](/deploy/vendo-cloud).