What it produces
Output from a real run, not a mockup. Two leads, both checked against their source before you ever see them.
npm run run -- "Find Jakarta startups that recently raised and might need AI research"
leads -> web_search {"query":"Jakarta startup funding 2025"}
leads -> fetch_url {"url":"https://east.vc/news/..."}
leads -> fetch_url {"url":"https://indonesiabusinesspost.com/..."}
checking every lead source in code...
VERIFIED Sxored
VERIFIED Rekosistem
| Leads verified / rejected | 2 / 0 |
| Human actions required | 0 |
| Model cost this run | $0.12 |
Every claim carries a source
The Leads agent must name a URL for each prospect. Afterwards, code opens that URL and checks the company is actually on the page. A lead whose source does not support it is rejected and shown to you as the agent's mistake, not quietly dropped.
It stops before anything consequential
Agents run freely until one reaches something that changes the world outside the workspace. That one stops and raises a card. The other seven keep working while you decide. Unanswered approvals expire and default to denied.
Honest numbers
“Human actions” and “leads verified” are real counts. “Hours saved” is an estimate from a fixed table, and the dashboard says so on screen. Every run reports what it actually cost, from the token counts the API returned.
Nothing to install
Zero npm dependencies — Node's own fetch, http and crypto. The whole thing is readable in an afternoon, which is the point: a stack you cannot audit is one you have to trust.