What an agent checks when it wakes.
Every agent wakes up in the past, frozen at its training cutoff. The briefing of what changed since: calendars, business days, dependency status.
Every Desvela product ends in a structured, signed event — not a dashboard. If you build agents, this page is the contract: how we wake yours, how you verify us, and how we behave on the open web.
Every agent wakes up in the past, frozen at its training cutoff. The briefing of what changed since: calendars, business days, dependency status.
Rules, capabilities and payment rails of any domain, in one lookup — before your agent touches it.
Deterministic watches, thresholds, signed webhook delivery. First product: AI Brand Watch.
Deterministic checks cost ~100× less than waking an LLM to poll. Your agent should spend tokens deciding, not staring.
{
"events": [{
"event": "brand_visibility_changed",
"brand": "Acme CRM",
"engine": "perplexity",
"query": "best CRM for small business in Spain",
"change": {
"type": "disappeared",
"share_of_voice": { "before": 0.67, "after": 0 },
"new_cited_source": "reddit.com/r/smallbusiness/..."
},
"checked_at": "2026-07-07T09:00:00Z"
}]
}
Signed with X-Desvela-Signature (HMAC-SHA256, secret issued on your first run). Verification snippet and full reference in the docs.
POST with retries and exponential backoff. Redirects refused. Fires any framework that can receive HTTP — LangGraph, OpenAI Agents SDK, n8n, a serverless function.
For episodic agents with no public endpoint: read the accumulated events on your next run. The watching already happened; you just read the summary.
For the human-agent hybrid: same structured event, rendered readable. No dashboard in between.
Our watchers identify themselves (Desvela-BrandWatch/0.1 (+https://desvela.dev/bot)), respect robots.txt by default, and never disguise themselves to defeat blocks. If a site does not want to be watched, we tell you instead of cheating — and we never charge for checks we could not run. This site ships an llms.txt, and desvela.dev publishes an ai-catalog.json manifest (ARD) — we preach the agent-readable web and practice it.
On the roadmap after launch: MCP server under the ai.desvela/* namespace (create and manage watches from Claude, Cursor or any MCP client) and x402 pay-per-call on the REST endpoint.