Product
One API between your agent and the live web.
Three endpoints (search, extract, monitor) over a partial corpus we crawl and curate ourselves (deepest in travel today), answered as structured JSON your model can reason over.
Search: ask in plain language
Your agent asks a question; SeaWeb returns ranked passages and structured entity cards from its own crawl and curated indexes, each one carrying the page it came from and the date it was seen.
Scope a query to destination sentiment or live disruption alerts. Generic search never falls back to an external provider.
Extract: public HTML as clean text
Hand over a public HTML URL and get the readable document back: title, clean text, passages. SeaWeb checks its index first, then attempts one live fetch on a miss. Robots, publisher directives, and private-network protections still apply.
Monitor: a standing watch
Register a query once and SeaWeb keeps evaluating it after your agent disconnects. When something changes, your webhook gets a signed POST. No polling loop to run, no schedule to babysit.
Every monitor has an end date you set. Stopping it removes its standing query; alerts inform and never act on your behalf.
An honesty label on every result
Most search APIs return something no matter what. SeaWeb grades its own answers:
every result carries match_quality,
and every response says whether the subject was actually covered, or marks the
rows uncertain so your agent treats
them as leads, not facts. Retrieval you can build guardrails on.
{
"coverage": "covered",
"results": [
{
"title": "Kyoto in autumn",
"url": "https://…/kyoto-autumn",
"fetched_at": "2026-07-31",
"match_quality": "strong"
}
]
}Bring your own data
The same engine that searches our web indexes can be pointed at yours. Teams run SeaWeb over internal knowledge (trace logs, docs, tickets, product databases) indexed alongside SeaWeb’s web corpus behind the same MCP tools, so one agent surface answers from both.
Retrieval is hybrid: keyword and semantic, routed per query. Cost doesn’t scale the way pure vector search does. We build each internal deployment to the company’s request.
Before / after the honest index
Local-service sites show photo galleries. For agents, the useful contrast is what happens when the corpus cannot answer: invent a lookalike, or abstain with a label.
Generic web search returns a confident paragraph even when the source set never covered the company, venue, or region you asked about.
SeaWeb returns structured rows with coverage / match_quality so the agent can discount or retry instead of acting on a hallucination.
Plugs into the agents you already run
SeaWeb speaks MCP, so Claude Code, Cursor, Codex, claude.ai, ChatGPT, and any MCP-capable client get the tools automatically. No SDK to install. Not MCP-aware? The gateway is plain HTTP JSON-RPC. Setup for every client is in the docs.
Claude Code, Cursor, Codex, claude.ai, ChatGPT, VS Code, Windsurf, and raw HTTP.