Frequently Asked Questions

Gmail’s filters have barely changed since it launched in 2004: you still hand-write rigid keyword rules, and its spam catching isn’t always reliable. Even with the Promotions and Social tabs, our users get all kinds of unwanted mail Gmail never manages properly, and they still spend precious minutes a day sorting their inbox and deleting junk they didn’t want. The deeper problem is that most unwanted mail isn’t spam; it’s legitimate mail you simply don’t want, and a keyword rule can’t express “kill the card-offer spam but keep my statements and fraud alerts.” Kelp reads each message the way a sharp assistant would, weighing who sent it and what it actually says.

Mostly, and Kelp leaves Gmail’s spam filter running. But a steady trickle gets through, especially targeted scams built to look legitimate: brand impersonation, fake invoices and DocuSign requests, “hi, it’s me” notes from throwaway accounts. Kelp reliably catches these, weighing sender authentication and what a message is actually trying to do the way a wary human would. It works as a second layer behind Gmail, not a replacement for it.

It’s more than that. A spam filter catches mass abuse and stops there; Kelp also works the much larger gray area of mail that’s perfectly legitimate and still unwanted: cold pitches, newsletters you’ve cooled on, mail meant for someone who happens to share your name. It catches the targeted scams that slip past Gmail too, but that’s the floor, not the ceiling. And you steer all of it in plain English.

Kelp isn’t a new inbox to migrate to or a writing assistant. You keep using Gmail or Superhuman exactly as you do now. It’s an autonomous filtering agent aimed at the cold mail you never wanted to see, and it improves itself: it notices the corrections you make and proposes its own changes, backtested against your real mail, for you to approve.

By design, plenty. It doesn’t write or send email, doesn’t read or touch mail from people you know, and won’t rearrange your mailbox at large. Its job is sorting cold inbound and managing the filters that do it. Today it works with Gmail and Google Workspace only, including clients like Superhuman that sit on top of Gmail; other providers aren’t supported yet.

Not yet. Today Kelp deliberately protects the people you know: mail from your contacts and trusted senders always reaches you, untouched, so it never risks hiding something that matters. Filtering that lane too is a bigger, more delicate problem, and we’d rather get it right than rush it. It’s an area we’re actively exploring and hope to open up soon.

Not yet. We’re focused on clearing out the new mail you get each week, so Kelp keeps your inbox from filling back up rather than digging through the backlog you’ve already accumulated. A one-time cleanup of a six-figure archive is a different project, and one we may tackle down the road.

No, and this is a point we care about. The AI never receives your Google login or an open connection to your mailbox; that stays inside Kelp’s own servers. It works only through a small, fixed set of read tools, so it sees a message through those tools, never by reaching into your mailbox itself. For automatic filtering it only ever sees cold mail from people you don’t know: mail from your contacts and the senders you trust is never read, and anyone you’ve blocked is removed before it reaches you. The AI can’t browse your mailbox at will, send, or rearrange anything outside that toolset.

No, and keeping as little of your data as possible is a deliberate design choice. Message bodies are held in memory only long enough to classify them and are never written to our database. We don’t keep copies of your inbox or your contacts: the contacts Kelp recognizes are stored as one-way hashes, not the addresses themselves. The one optional exception is a 30-day audit trail of the mail Kelp filters (the sender, subject, and first few words) so you can review its work, and turning it off in Settings purges it immediately.

Read the full security details →

Never. Your mail doesn’t train any model (ours or anyone else’s), and it is never sold or shared. The AI we use to read messages doesn’t train on it either: data we send through Anthropic’s API isn’t used to train their models. Kelp earns money from subscriptions, so you’re the customer, not the product.

It largely doesn’t. Gmail bundles those permissions, so the screen lists them even for apps that don’t use them. Kelp has no sending code at all; it only sorts what arrives. The strongest thing it ever does is move a message to Spam, Trash, or your archive, all of which you can reverse while the message is still in Gmail.

A fair question to ask anyone. Kelp passed Google’s CASA security assessment (run by accredited third-party assessors) before Google granted production access to Gmail, so it’s held to that bar. We store as little of your data as possible, our revenue comes from subscriptions rather than data, and every action is transparent and reversible. Disconnect from your Google account whenever you like and Kelp stops immediately.

Kelp for security teams →

Kelp can move mail to Spam or Trash, so like any filter (Gmail’s own keyword filters included) there’s a small chance it tucks away something you wanted, and if you never check those buckets Gmail eventually clears them. We work hard to avoid that: when Kelp isn’t sure it leaves the message where it is, it never touches mail from people you know, it shows you exactly what a new filter would catch before you turn it on, and it learns from the corrections you make. For extra peace of mind you can choose gentler actions, like label-and-archive (the message leaves your inbox but stays in Gmail, fully searchable) or a delayed delete that holds mail for a grace period first. Every action is also recorded on your dashboard and reversible while the message is still in Gmail.

Every action Kelp takes is recorded on your dashboard and reversible while the message is still in Gmail: re-inbox it, drop a label, or pull it back from Spam or Trash, and blocks are reversible too. The one way mail can genuinely be lost is the same as with any filter, Gmail’s included: a message sent to Spam or Trash that sits there past Gmail’s retention window without anyone noticing. The gentler options (label-and-archive, which leaves the message in place, or a delayed delete that holds it for a grace period) exist to widen that margin.

Every message Kelp acts on is recorded with the plain-language reasoning behind the decision, so you can always see why it was filtered. And everything Kelp has learned about you (your work, who matters, what you’ve declined) sits on one page you can read, edit, or delete. It learns quietly, never secretly.

No, and that’s by design, not by hoping the AI behaves. When Kelp reads a cold message, the only thing it can decide is whether that message matches your filters; it can’t grant a sender trust, change your settings, or follow instructions hidden in an email. Trust comes from your real contacts and the senders you’ve chosen, and Kelp checks sender authentication so no one can impersonate them. The worst a manipulative email can do is fail to match a filter and stay in your inbox.

A straightforward subscription. We don’t sell data, run ads, or monetize your inbox in any way; the only thing we sell is the product. That keeps our incentive aligned with yours: a quieter inbox, not more time spent in it.

Every account starts with a 14-day free trial. We’re running special launch pricing for early adopters right now, and the current rates are always on the Pricing page.

See launch pricing →