Every tool answers with text for the model to read. Where there is more — a plan id, a run record, candidates — it is also in the result’s
structuredContent. Refusals (“this needs a goal”) are ordinary answers, not errors: read them and call again.
Look around
orbbai_who_knows_them
One person in; the people around them, each with how strong the tie is and the evidence under it — who they worked alongside and for how long, who they interact with publicly, and where each of those people is now. One call: no plan, no approval, nothing to poll.
It pins the person itself. A LinkedIn URL or public id is read directly; a full name is resolved on the open record, and the answer says how sure it was. Where a name really is several people, it lists them instead of guessing.
A LinkedIn profile URL, the public id after
/in/, or a full name. A single word is read as a public id, never as a name — for somebody known only as “Dana”, send the full name.What the introduction is for, if the question says —
"revenue", "engineering", "anyone senior". Ranks and filters the people around them by what each does now and did then. Without it, the answer is whoever they know best rather than whoever matters here.Your own user’s LinkedIn URL. The answer then says which of these people your user has worked alongside themselves — a path that needs no introduction.
This reads a few people off the record and cannot rank them against a goal. “Every relationship, scored and ranked” is a research run.
orbbai_who_knows_someone_there
One company in; ways in out. Each is a named person who left that company and a named person still inside whose time there overlapped theirs, with how many months they were there together. The person who left is the one you can approach.
A domain (
snowflake.com), a name, or a LinkedIn id company/<slug>. A name is a guess, and the answer says which company it read. A LinkedIn company URL is not read — send the company/<slug> id.A city, state or country, when the question names one. Somebody whose record names no place is not excluded: “we do not know where they are” is not “they are somewhere else”.
Orient
orbbai_capabilities
What Orbb can and cannot answer. Called bare, a short overview and the questions it settles, with how much of each is public record and how much needs data only you hold.
overview (the default), kinds (the kinds of research, their limits and depths), asking (what makes an introduction land), questions, or all — every part at once, and long.orbbai_playbooks
A library of worked workflows: scoring accounts by who you can reach, finding the route into every open opportunity, spotting deals with no relationship behind them, prioritising outbound by warm access, finding a missing executive sponsor, picking an event invite list. Each says when it runs, the steps, the calls it makes, and what your own systems must supply — either to run once now with these tools, or as a build spec for your own automation.
What your user is trying to do, in their words —
"prioritise outbound", "we keep losing deals we have no relationship into". Returns the two or three that fit. Omitted, the shortest overview.orbbai_held
What your account has already researched, by kind and date. An answer from research already done is immediate and free; researching the same subject again spends tokens to learn nothing new.
Research
orbbai_people_resolve
Turns names into the handles orbbai_plan takes. Returns candidates and never picks: a common name comes back as several people, each with a headline and company.
1 to 25 names. Each is a full name or a LinkedIn URL. Add a company or title where you have one —
"Dana Mor, Head of Data at Acme" resolves far more often.What this is for, in your user’s words. Later calls in this conversation weigh answers against it.
resolved has the name asked and its candidates — handle, name, headline, company, and held when Orbb already has research on them. settled is set when exactly one researchable profile matched.
orbbai_plan
What a research run would take: a price in Research Tokens, a time, and a planId. Nothing starts. What you give decides the kind of run, and every kind takes the goal in ask.
What your user wants, in their own words — a sentence, not a filter. Optional in the schema, but most plans are refused without it.
One company, by name or domain. Plans a company run: who inside it is worth going through. Answers with a
run id for the shortlist rather than a planId.Up to 20 roles —
["VP Data", "Head of Analytics"]. With country or cities, plans a search across the whole public index, with a column per condition and a yes, no or not-checked in every cell.planId (or run for a company), questions worth putting to your user, an optional outside block, and refused when the plan cannot run — with the one thing missing.
orbbai_run_start
Agrees to a plan and starts it. Spends Research Tokens against the account the server is signed in as — the plan carries the price. Takes the plan id and nothing else.
The structured result carries the run id, a link to it in the console, and reused: true when a run for the same subject was already working.
orbbai_run_status
Where a run is and what it has found so far. Runs take minutes; poll.
orbbai_run_stop
Stops a run and keeps what it found — the rows stay readable through orbbai_run_status. No refund for what was spent. Harmless on a run already finished.
Use the results
orbbai_routes_from_research
Routes out of research already finished, ranked against your goal, with the reasoning and two warnings about how to read them. It reads only what your account has already researched, and researches nobody: on a subject nothing has been run on, it says so.
What your user is trying to achieve. Required here — without a goal there is nothing to rank against, so it refuses.
orbbai_table_narrow
A new table from one a run already built, keeping the rows that meet conditions written as sentences. The original is untouched. Each row is judged against the conditions, at a small cost per row, so send every condition in one call.
orbbai_ours_remember
Records who is on your own side — colleagues, advisors, investors, customers, former teammates. Every later answer takes them into account, and routes through them rank first. Adds and never removes.
Your user’s own word for the relationship —
"design partner", "former colleague", "board member". The vocabulary is open on purpose.