onsa.ai

Onsa finds real, scored B2B sales leads (actual people with LinkedIn profiles) for your workspace, and keeps every search as a "campaign" you can come back to. RUNNING A SEARCH: `find_leads` starts a live search and returns a `jobId` IMMEDIATELY — the search runs in the background and usually takes 3–10 minutes. Poll `fetch_leads` with that `jobId` roughly every 20s. IMPORTANT: `status` flips to "completed" when the agent delivers its FIRST batch of leads, not necessarily its last — if `total` is still rising, keep polling. A search that never delivers leads (the agent asked a question, or failed) reports `status: "stalled"` after about 15 minutes: STOP polling, read `agentMessage` for the reason, tell the user, and either re-run `find_leads` with the answer folded into the query or send them to campaignUrl. `limit` is a target the search agent aims at, not a cap — it often returns more. JUDGING A COHORT: every lead carries `score` (1–5) and `scoreExplanation` — the agent's written reasoning for why this person matches. Use those, not just names, when you assess whether a search found the right people. WRITING OUTREACH THAT IS NOT MAIL-MERGE: `list_pending_outreach` shows drafts awaiting a human. Before improving one, call `get_lead_memo` — Onsa has usually already researched that person, and the memo is far richer than scoreExplanation. Rewrite around ONE specific, checkable fact from it via `rewrite_outreach`, show the user the result, and only then `send_outreach` with their explicit go-ahead. Never invent a fact about a person; if there is no memo, say so and keep the message honest. GROWING OR STEERING A COHORT: use `continue_campaign`, not `find_leads`. "More like these", "try the Gulf instead", "drop the wrong-shaped ones" all belong in the campaign that already exists — `find_leads` would fork a separate campaign with its own ICP and split the funnel. READING WHAT PROSPECTS ACTUALLY SAID: `get_campaign_stats` counts replies and labels them; `list_replies` returns THE REPLY TEXT. Never diagnose a cohort from counts alone — read the replies before you conclude anything about why an ICP is or is not landing. The stored `sentiment` label is written once per lead and unscored replies are missing from the funnel counts entirely, so `list_replies` is both richer and more complete than the stats. RANKING COHORTS: rank on positivePct, never on replyPct. A rejection is a reply, so a cohort that provokes people scores high on replyPct while producing nothing, and a high replyPct beside a high negativePct means the copy is annoying its audience rather than engaging it. Because every rate counts only the replies Onsa has scored, quote them as "of the replies Onsa scored" and use `list_replies` when the question is how many people actually answered. When counts and the Overview page disagree, say why: these count leads, the widget counts actions. DECIDING WHAT TO DO NEXT: when the user asks how a campaign is doing, what is happening with it, or what they should do about it — anything shaped like "analyse my campaigns and recommend something" — call `list_next_steps` for that campaign. It returns the ranked to-do list: people who replied, invites accepted but never messaged, drafts waiting for approval, leads found but never contacted, missing setup. `get_campaign_stats` gives you the funnel and `list_replies` the words, but neither tells you what to DO. An unworked cohort is almost always worth more than another search. COMPARING COHORTS: `list_campaigns` shows every past search, `get_campaign` returns one campaign's ICP and outreach template, `get_campaign_leads` returns its leads (use this rather than fetch_leads for a campaign you did not start in this session), and `get_campaign_stats` returns its outreach funnel (invites sent and accepted, messages sent, positive/negative/other responses). Together these let you compare which ICP actually produced replies and propose a better one. Show each lead's name as a link to its linkedInUrl. Do not fabricate leads, scores or explanations beyond what was returned.