One entry per ship. Dated, plain, no adjectives.

Last release opened the agent channel to your customers. Two things about getting connected were wrong, and both are fixed.

The setup we handed you now includes your key

Every ready-made setup — Claude Code, Cursor, VS Code, Codex, Windsurf — gave you the address and left out the credential, so pasting it in was refused. They were written back when the address was all you needed. Each one now carries your key, and the Agents page shows a working example next to your endpoint, with a link to where your key lives.

If connecting an assistant failed for you before, the setups on that page will work now.

/agent only appears when it will actually work

The command your customers use to connect appeared for people it couldn't help — before you'd sent us anything to identify them by, or on a plan or setting where connecting isn't on. They'd follow the steps and be turned away at the end.

It now appears only for someone who can genuinely connect right then. And when it does appear, it hands them a ready-made setup with their own address already filled in and one thing to replace — their own key for your product — rather than describing what to type.

They bring their own assistant, and their own key

Someone signed into your product can now point Claude, ChatGPT, or any other assistant at their own account — to ask questions, get things done, or reach a person. They authenticate with the API token they already have for your product, so there is no key for you to hand out and nobody waiting on your approval.

They find their address by typing /agent in the assistant on your site. It shows the address, points them at the page in your product where their token lives, and shows exactly how to send it. It appears only for people you've identified — visitors never see it.

You decide how customers connect

Agents → Policy now carries the choice:

• Straight away — their own token connects them (default)
• You approve — only a key you issue works
• Not allowed — nobody outside your team connects

Change it whenever you like; it takes effect on the very next connection.

Connecting is properly documented

Every guide — Claude, ChatGPT, Grok, Perplexity, plus Zapier, Make and n8n — now says exactly which credential each address takes and where to find it. If you tried connecting an assistant before and were turned away, it's worth another go: those guides used to say the main address needed no credential at all, which was wrong.

Assistants tell people what actually happened

The instructions a connected assistant receives now spell out which requests it carries out itself and which it only walks someone through — so it can't report a change as made when it wasn't.

"What Peeve never reads" now actually reads nothing

Guardrails lets you list fields Peeve should never look at — a card number, an account balance, anything you'd rather it didn't see. That list wasn't being applied. It is now: the fields you list are stripped from what Peeve reads before it leaves the browser, and it will not type into them.

Passwords are excluded whether you list them or not.

Your customers' own actions

If you've connected your own API or tools, Peeve can now offer those to a customer over email and social too — not just on your site. It always shows exactly what it's about to do and waits for the customer to agree.

Faster, and cheaper to run

Peeve now reuses work it has already done rather than repeating it on every message. You'll notice slightly quicker replies; we notice the bill.

Also

• Removed the Gorgias importer. It was listed but never finished, and we don't serve e-commerce yet — better to not offer it than to fail halfway through a migration.

Your agent can now do things over email and social

Until now Peeve could answer a customer on email, WhatsApp, Messenger and Instagram — but anything that meant changing something went to a person. Now it can handle refunds, cancellations, plan changes and your own connected actions on those channels too.

Nothing happens without the customer agreeing first. Peeve sets out exactly what will change and waits for them to reply — and only a clear yes counts. A question, a maybe, or "yes, but…" leaves everything untouched.

Your customers' AI assistants can talk to your agent

A customer who connects their own assistant can now hold a real conversation with your product: ask questions answered from your Brain, and reach your team when they need a person. Those conversations show up in Sessions and Hand-offs like any other.

Anything you've marked as a person's job becomes a hand-off rather than a dead end — the customer still gets an answer, it just comes from you.

Your own actions, not just billing

If you've connected your own API or tools, Peeve can now offer those to customers the same way — with the same confirmation step, and showing exactly what it's about to send.

Clearer control over what your agent may do

Agent policy is now split by who's connected — your team's assistants and a customer's own assistant are different situations, and the same setting meant different things in each. Each now says plainly what it does.

"What only a person may do" now applies everywhere — on your site and to anything Peeve does through a connected service.

Notifications that actually arrive

Five alerts in Settings → Notifications were switchable but never sent. They work now — including New hand-off, so you're told when a customer asks for a person even if you haven't connected Slack.

Fixes

• The agent's own replies were missing from the conversation view on WhatsApp, Messenger, Instagram and API conversations — teammates saw only the customer's side.
• Usage broke social conversations out incorrectly, counting them as live chat.
• Analytics counted email and social conversations as website visitors.
• Sessions can now be filtered by every channel; only two of six were listed.
• Turning "Session cap" off didn't stop those alerts.

Peeve is an AI customer support & experience agent that sits on top of your product. It doesn't just answer questions — it walks your users to exactly where a feature lives, helps them get things done, captures and qualifies leads, books meetings, and hands off to a human the moment it should. Everything it says is grounded in a verified Brain built from your own product and knowledge, so it never guesses.

Here's what's in the box today.

The Brain — grounded, not guessing

• Automatic setup from your site. Point Peeve at your product and knowledge; it crawls both and builds a verified graph of your real capabilities (things a user can do — "invite a teammate", "download an invoice", "change plan") and the routes where each one lives.
• Verified before it's used. Only confirmed, allowed capabilities are ever exposed — so answers and actions come from your product's real truth.
• Stays fresh. Sources re-crawl on a schedule; a review queue and conflict resolution keep the Brain accurate as your product changes.
• Import what you already have from Intercom, Zendesk, Voiceflow and more.

One agent, four ways to resolve

Every request resolves down the right path, automatically:

• Answer — reply from your product's grounded knowledge.
• Show — walk the user to exactly where a capability lives (Peeve is show-first).
• Do — take an action with the user, with confirmation — never a silent write.
• Escalate — hand off to a human when that's the right call.

Everywhere your users are

Peeve works across every channel from one Brain and one credit pool:

• On-site widget — the cursor that guides and answers in-product.
• Email, WhatsApp, Messenger, Instagram, Slack, and voice — AI answers first, grounded in your Brain, and escalates only when it should.
• Agent channel (MCP) — your users' own AI assistants (Claude, ChatGPT, or their own) connect to your product through Peeve and read the same Brain, with the same permissions — never more.

Humans, when they're needed

• Hand-off inbox — a clean queue of conversations that need a person, across live chat, email, and social.
• Routing rules & saved replies, plus assignment and send-to-helpdesk.
• Helpdesk integrations — push a hand-off straight to Zendesk, HubSpot, or Intercom.

Know your users, grow your pipeline

• Contacts — every user and visitor, their history, and who's getting stuck.
• Leads — captured and qualified from real conversations.
• Meetings — Peeve pre-qualifies prospects with a short interactive form, then offers a booking through Cal.com or Calendly.

Connect your stack

Bring your tools into the agent: Stripe, Salesforce, Cal.com, Calendly, HubSpot, Zendesk, Intercom, Slack, plus your own API or MCP server — so Peeve can act where your product and business already live.

See what it's doing for you

• Analytics — resolution rate, self-serve rate, routes and paths, channels, conversion, NPS, and the support cost you've saved, with CSV export.
• Sessions — a full, replayable record of every conversation on every channel.

Controls you can trust

• Guardrails & a kill switch — stop Peeve instantly, set write policy, and keep agents to the same permissions as your users.
• Roles, 2FA, and audit logs for your team.
• Multiple workspaces from one account.

Billing that's simple

• One universal credit pool for every channel — one credit is a conversation on any channel, 2 voice minutes, or 100 email scans. Clear usage, plan upgrades, and instant top-ups.

Made to look right

• A refined, glass interface — and Light & Dark themes that follow your device or lock to your preference, per teammate.

This is just the beginning. We'll ship improvements continuously — watch this space.