The grounding engine
for AI agents.
An agent that acts on a wrong fact is worse than one that never answered. Ipnops doesn't return ten links to read — it returns the answer as data: claims, evidence quotes, calibrated confidence. Grounded on a live-web index, verified before it reaches your model.
01 / 06
of web traffic is now automated — agents crossed the majority in 2026, 18 months ahead of forecast.
Cloudflare Radar, 2026
faster than humans: agentic AI traffic grew roughly 7,851% year over year.
HUMAN Security, 2026
of bot activity now targets APIs directly — machines skipping the human UI entirely.
Imperva Bad Bot Report, 2026
02 / 06
An engine that reads the web, so your agent doesn't have to guess.
Retrieval built for people fails agents in a specific way: hallucinated citations, dead links, and confident answers when the source came back empty. Ipnops closes that gap at retrieval time — every answer is decomposed into claims, every claim is verified against the passage it cites, and when the evidence isn't enough, the response says so rather than guess.
One live-web index
The index stores structured passage chunks with claim provenance — hybrid-retrieved and reranked, kept current as the web moves — so every answer traces to a passage rather than a guess.
Claims, not prose
Every assertion arrives as data: the claim, its verdict, the literal quote behind it, and one calibrated confidence for the whole answer. An agent decides to trust, verify, or escalate — without re-parsing prose.
Grounded or it abstains
A verification pass grades every claim against its cited passage. Conflicting sources are surfaced and attributed, never averaged. And when the evidence doesn't cover the question, the answer reports insufficient instead of improvising.
03 / 06
One question. Every way in.
A grounded answer shouldn't depend on where the fact lives. Ipnops runs a ladder of reach — the live-web index, place data, and, when there's no endpoint at all, operating the interface directly. Every rung is grounding-checked, so a miss degrades cleanly to the next instead of guessing.
Live-web index → place data
A question resolves against the live-web index, grounding-checked; anything spatial resolves through OSM/Nominatim place data. Each tier is checked, so a miss degrades cleanly to the next rung instead of guessing.
When there's nothing to call, it operates
The data an agent needs is increasingly behind a login, on a canvas, or inside an app with no API. Ipnops drives a live browser — a plan → act → reflect loop — and, when the DOM lies, works the pixels with computer-use.
Ground once, replay forever
Grounding a fact usually means running the whole loop again on every request. Ipnops grounds a trajectory once, then replays it deterministically — turning O(steps × runs) model calls into O(1) on repeat work, so a warm path costs almost nothing to keep fresh.
04 / 06
Human search ends at the results page. Agent search is just getting started.
Classic engines live by one contract: answer in milliseconds or die. Agents don't need fast guesses — they need finished work. Ipnops runs searches that outlive the request: deep research that crawls what it's missing, and standing queries that watch a fact and call back when it changes.
Deep research as a durable run
One POST returns a run id. The engine answers, finds the claims it couldn't ground, crawls the gap, and re-answers — iterating until the answer holds or the budget is spent. Poll it, stream progress over SSE, or take a webhook when it closes. The run survives restarts; the index keeps what it learned.
Standing queries that watch the facts
Register a question once. Ipnops re-answers on schedule and diffs the grounded claims — a deterministic set comparison, zero extra model calls. Your webhook fires when the facts change, not when the phrasing does. Rewording is silence; a new number is a signal.
05 / 06
An MCP server, or plain HTTP. One line either way.
Point an MCP-native runtime at Ipnops and grounded search, answers, deep research, and monitors appear as typed tools — no SDK, no glue. Not on MCP yet? The same engine answers over plain HTTP. Either way the payload is model-shaped on arrival.
# one connector — Ipnops becomes an agent tool
{
"ipnops": {
"transport": "http",
"url": "https://api.ipnops.ai/mcp"
}
}# same call an agent makes over MCP — here as raw HTTP
curl -sS https://api.ipnops.ai/answer \
-H "content-type: application/json" \
-d '{ "query": "who maintains the sqlite file format?" }'
# ◂ returns the cited AnswerResult below — no parsing, no scrapingHybrid retrieve → rerank → typed cited highlights.
Lexical and dense semantic arms over the live-web index, fused and reranked. Returns passage highlights, each self-citing its source.
{ "highlights": [
{ "text": "…passage…",
"score": 0.91,
"source": { "url": "…" } }
],
"arms": { "lexical": true, "dense": true, "reranked": true }
}One grounded answer as typed data.
Consolidates highlights into a single answer: marker-per-claim citations, verified claims with evidence quotes, calibrated confidence, and a sufficiency verdict — the shape a model can act on without re-parsing.
{ "answer": "… [1] … [2]",
"citations": [{ "marker": 1, "url": "…" }],
"claims": [{ "text": "…", "verdict": "supported", "quotes": ["…"] }],
"confidence": 0.84, "sufficiency": "supported",
"grounding": "own-index"
}Deep research as an async run.
Returns a run id immediately. The engine iterates — answer, find the ungrounded claims, crawl the gap, re-answer — and checkpoints every step. Poll, stream SSE progress, or take the webhook.
{ "run_id": "run_5c1f…", "status": "queued" }
# GET /research/:id/events — SSE, resumes from Last-Event-ID
event: iteration → gap → crawl → completedA standing query with a webhook.
Watch a question instead of asking it once. Each sweep re-answers and diffs the grounded claim set deterministically; the webhook carries exactly what changed — added claims, removed claims, confidence delta.
{ "monitor_id": "mon_9b2e…",
"diff": { "addedClaims": ["the record is 105"],
"removedClaims": ["the record is 100"] },
"answer": "… [1]", "confidence": 0.81
}The MCP server itself — Streamable HTTP, stateless.
Search, answer, extract, research, and monitor as typed tools, each with an outputSchema. This is the front door most agents use.
{ "tools": [
"ipnops_search", "ipnops_answer", "ipnops_extract",
"ipnops_research", "ipnops_research_status", "ipnops_monitor"
] }
# each with an outputSchema — results arrive typed06 / 06
Ground AI on data that never leaves your walls.
Ipnops is a managed service. For enterprises and sovereign data owners — teams whose corpus can't cross a border, a VPC, or a compliance line — the same engine ships as an enterprise deployment inside your own environment. Retrieval, verification, and synthesis run where the data lives. Not retrieval bolted onto a chatbot: a grounding layer between your data and every model and agent you run.
The stack is yours to run
The full engine — index, hybrid retrieval, reranking, verification gate — deployed in your own environment against your own data. Nothing leaves the perimeter to be grounded.
Bring your own reasoner
Point synthesis at your own model endpoint — a hosted model, a private inference cluster, whatever your compliance requires. Paid egress for open-web reach is opt-in, not baked in: a fully-local deployment makes zero outbound calls.
Beyond retrieval-and-hope
Simple RAG retrieves text and hopes the model uses it well. The Ipnops layer types it: claims, verdicts, evidence quotes, confidence, freshness — an auditable trail from every answer back to the exact passage that grounded it.