Evidence infrastructure
for production AI agents.

Search retrieves things. unlob tells an agent what evidence exists, whether the sources are independent, whether the set is sufficient, what is missing, how fresh it is, and what to do next — within a token budget. One call, ground, returns that structure. No generated answer is returned, deliberately: the reasoning stays with whichever model you run, and the layer beneath it is deterministic.

The call says so. ground returns insufficient when evidence exists but does not meet the bar you set, stale when nothing is inside your freshness window, partial when part of the index was unreachable, and empty when nothing was found — each with next_actions saying what would change it. A model that reads the status can abstain correctly instead of answering from whatever came back, and that is the behaviour our agent benchmark measures first.

Because that puts a retrieval company in a race to run the most expensive reasoning on its own models. Exa, Parallel, Tavily and Linkup are all moving that way. unlob is the infrastructure their agents — and yours — ought to consume: the smallest defensible evidence package, how its pieces relate, and what is not known, beneath arbitrary reasoning models. It is cheaper for the same reason: the structure is precomputed, so a ground call costs infrastructure money rather than research-agent money.

Not by counting hosts. Within a story, passages are grouped into origins: a wire dispatch and its reprints are one origin, a press release and the sites that ran it verbatim are one origin, and two outlets under one media group are two origins with one owner. independent_origins counts distinct owners. Every label carries a confidence, and the rule is documented in full — seventeen URLs and two genuinely independent origins are different findings, and the receipt says which you have.

Every ground answer carries one: what was searched, what was not — shard slots down, verticals excluded, stories not expanded — how fresh the evidence is against the window you asked for, which source classes are present and missing, and a plain-language list of known gaps. It is why_not generalised from one URL to a whole request, so an agent knows when an answer is incomplete instead of inferring absence from a short list.

Our own. We crawl the open web ourselves and serve from the index we built, which is what makes the evidence graph possible: story clusters, origins, the link graph and corroboration signals are computed when the index is built. No upstream search engine can deprecate us, reprice us or rate-limit us.

Yes, over Streamable HTTP at https://api.unlob.com/mcp with your API key. The grounding profile (?profile=grounding) lists ground and get_document — the right surface for most tasks and far less schema in every request. The default profile lists all 13 tools, including the six coverage-graph operations. There is also an installable Agent Skill at api.unlob.com/skill.md, so a coding agent can learn the API without a human reading the docs.

10,000 credits a month, 30 days of history, and 60 credits a minute — no card required, and it is the most generous free tier in the category. A search costs 1 credit and a ground call 5 credits, however many stories it expands. The rate limit, not the allowance, is what makes it a tier for building and evaluating rather than for serving users; it hard-caps at its monthly credits rather than billing you by surprise.