Relay — Autonomous SDR

The complete source: the nine-stage agent loop, the evidence gate, the contact finder, the compliance layer, the React operator console, the database migrations and 115 tests — plus an install guide and a licence for your tier. It deploys to your own Cloudflare account (Workers + D1) and calls Anthropic under your own API key. Nothing routes through us, and there is no usage meter on top. There is also a $0 lane: `npm run dev:free` runs the entire loop against a real model through your local Claude CLI, so you can watch it research and draft before you have spent anything at all.

Those are good at volume, lists and deliverability plumbing. Relay is not trying to win on volume — it sends fewer messages, more slowly, and refuses some of the ones it wrote. The difference is the gate. A model asked to 'personalise using this research' will happily produce a sentence the research doesn't support, and every tool in this category ships that sentence. Relay's verifier is 420 lines of ordinary code with no model in it: a claim must cite a source Relay itself fetched, the quote must appear verbatim in that source, and every particular in the claim — every figure, every month, every funding stage — must appear in the quote. If it doesn't, the message stops and tells you which sentence failed and why.

Proven: 115 automated tests, 29 of them on the gate; a real-model verification pass through a live Opus model; and one real end-to-end run that hunted 53 companies, disqualified 30, found 10 contacts and produced 8 verified drafts — while refusing 7 others. Five genuine defects were found in the gate by running it rather than reading it, and each is now a regression test. Not proven, and we would rather you heard it here: Relay has never been deployed to production by us — every run so far has been on the local runtime, which executes the same routes and the same SQL, but a first `wrangler deploy` is still a first. And no email has ever been sent by Relay itself; the drafts from that run were handed to a human's mailbox, which is what it does below the top autopilot level. The full proven/unproven ledger ships with the kit in docs/VALIDATION.md — it is written so you can tell the difference yourself.

It does not automate LinkedIn — that would put your account, not ours, against their terms. It does not buy contact data. It does not send anything at all below autopilot level 2, and even at the top level it will not release a draft the gate refused; there is no override button, because an override button is how this product stops being this product. And it makes no deliverability promises: the warmup ramp and sending caps are conventional practice, and your domain reputation is yours.

To deploy it, yes — you need to be comfortable with a terminal, a Cloudflare account and five `wrangler secret put` commands. Budget about half an hour with the install guide. Once it is running, the console is the product and anyone on your team can work the approval queue. If you would rather not run it yourself, reply to any email on this site and ask about the managed version.

No, and it would be easy to imply otherwise. Relay came from a working prototype rather than from our research engine, and it has not been through the demand gate our developer kits are scored against — so nothing here claims market evidence we do not have. What we can show you is the code and the run: the tests, the gate, the 53-company hunt, and an honest list of what has never been exercised. Nobody has paid for this yet. That is the part only you can change.