The evidence layer for agents
One call, ground, turns an objective into the smallest defensible
evidence set: who originally said it, who republished it, whether the sources are
independent, how fresh it is — and a receipt that says what was searched, what was
not, and whether the evidence is enough. No generated answer; the reasoning is
yours. Underneath sits an admission-controlled web index, a graph over it, and a
search API that returns passages rather than links.
Start here
ground
Evidence with provenance and a coverage receipt, packed to a token budget. Read
status first.
REST
23 endpoints
One GET per question. Generated from the API's own description.
MCP
13 tools
Two on the grounding profile, all of them on the default. Same parameters as REST.
Four things worth knowing first
Authentication
One key, two header spellings, and the minute after you mint it.
Search modes
Keyword, semantic, hybrid — and the query grammar.
Filters and sorting
Every filter, what it narrows, and which ones only bias.
Reading a result
What a hit carries, and the one field that changes what it means.
Still deciding?
These pages assume you have chosen unlob and want to use it. The case for choosing it is on the main site: the 2026 landscape maps every web search API with sourced pricing, the comparisons each name where the competitor wins, and why an own index matters covers what you inherit when a search API resells someone else's results.
For agents
Everything on this site is also readable by machine, without a key: skill.md (the Agent Skill), llms.txt, openapi.json, mcp/tools.json, and describe, which reports what this deployment can do right now rather than what the docs said last month.