How it works

Call Lens transcribes and analyses sales calls. An integration is a loop with four moves, and it is the same loop every time.

You push; we push back

There is no polling step. Everything Call Lens produces for a call arrives on one signed webhook, once, when processing finishes. If you find yourself writing a GET to check whether a call is ready, the integration has gone wrong somewhere earlier.

There is a way to read analysed data back on demand — the MCP server, which AI clients connect to — but it is a separate surface with its own credential, and it is not part of this loop. Nothing in an integration should wait on it.

That has a consequence worth designing for up front: your webhook endpoint is not optional plumbing, it is the only way you receive results. Register it before you submit anything.

The audio never passes through the API

POST /uploads hands back a URL and a ticket. You send the bytes straight to that URL, and only the ticket goes to POST /calls. Call Lens never proxies the recording.

Two things follow. Apply the upload.headers from the presign response to your PUT verbatim — the map is empty against some storage backends and required by others, so skipping it works in development and fails in production with a signature mismatch. And size the upload step for your own network, not for the API's rate limit.

Accepted is not analysed

POST /calls returns 201 the moment the call is queued. It is an acknowledgement that Call Lens has taken responsibility for the recording, not a result. The analysis arrives later, on the webhook.

200 instead of 201 means you have submitted this call before and Call Lens returned the original rather than making a duplicate. See idempotency.

What you are responsible for

Next: the quickstart walks the whole loop with real requests.