Say what you sell
Paste a URL or describe the offer. The agent reads your product, works out the value proposition, and proposes an ICP — the kind of company that has the problem you fix. Already know your ICP? Give it that instead.
Give it the URL of what you sell. It works out who buys it, then writes the code to go and find them — sourcing, enrichment, qualification — tests that code against a contract, runs it every day, and repairs itself when a source changes underneath it.
Try it nowSources
Every other tool hands you a list of connectors and asks you to work within it. Here there is no list. If a source has a page or an API, the agent writes the block that reads it, and writes the tests that prove it still works.
A few of the sources a pipeline has been built against.
How it works
Paste a URL or describe the offer. The agent reads your product, works out the value proposition, and proposes an ICP — the kind of company that has the problem you fix. Already know your ICP? Give it that instead.
Sourcing, filtering, enrichment, qualification, message. Each step becomes a block with an explicit data contract, a real implementation, and a test suite. Blocks run in an isolated sandbox until they satisfy the contract.
Every run is checked against the contract and against its own history. When a source changes shape, an agent gets the code, the logs and the failing rows, and ships a fix as a new version.
Anatomy of a block
Not a prompt fired at a model for every row. A versioned implementation with an input contract, an output contract, a test suite and a run history — so you can open any step and see exactly how a result was produced.
headers = fetch(website).headers
detected = "cf-ray" in headers
return {
"detected": detected,
"confidence": 0.99 if detected else 0.4,
"evidence": [h for h in CF_HINTS if h in headers],
}Self-healing
Nobody writes a scraper that works forever. The difference here is what happens on the morning it stops: the contract catches it before the data reaches your CRM, and an agent is already on it.
The pipeline executes on schedule and produces rows.
Rows are tested against the contract and compared with history — count, shape, fill rate, types.
Product Hunt returns zero launches, or the email fill rate halves. The run is held, not shipped.
An agent receives the current code, the logs, the failing rows and the historical results.
It rewrites the block, updates the tests, and runs the suite in the sandbox.
Green, and the new version goes live. Red, and it stays quarantined while the last good version keeps running.
What comes out
Not a list of company names. Each row carries why it matched, who to reach, what the agent found, and a message written from that context — plus a link back to the run that produced it.
Free while it is in beta. Sign in with Google, give it a URL, and watch it build the first pipeline.