Reads are automatic
Every write is confirmed and allow-listed. Nothing changes in your systems without approval.
AI Product Support Agent
Show, don't tell.
Peeve is an AI customer support agent that resolves questions inside your product. Where a chat widget describes the answer, Peeve highlights it, then clicks it, in your user's own session.
Installs in an afternoon · From $40/mo · No sales call
routed: SHOW · 96ms · confidence 0.94 · app.peeve.aiyou can take over any time
Peeve is the AI Product Support Agent for SaaS teams. It resolves support inside your product (answering, guiding users through the UI, completing safe confirmed actions, and escalating with full context), so users finish the task instead of reading instructions.
Your help center was accurate eight months ago.
Your product shipped 14 times since. The docs didn't. So your chatbot quotes a page describing a button that moved, your user asks a human, and someone on your team answers the same question for the fortieth time this month.
Peeve doesn't quote the page. It reads your live interface, finds the thing, and takes the user there.
help-center/billing/change-plan
1. Open Settings
2. Click Account & Plan
3. Choose your new tier
broken · renamed 4 months ago
what Peeve sees instead
/settings/billing → “Plan & usage”
verified 6 minutes ago · confidence 0.94
The Peeve Ladder
Peeve climbs only as far as it needs, and hands off before it guesses.
Four rungs of effort: Answer, Show, Do, Escalate. Peeve starts with the lightest touch and steps up a rung only when the question demands it, deciding which one in about 150 milliseconds. The full ladder runs inside your product; on your marketing site and docs, Peeve answers and hands off. One brain powers both.
01
the first rung
Answer
The lightest touch: returns the answer with its source attached.
“what's included in Pro?”
Answer →02
↑ climbs when the answer is a place, not a fact
Show
Peeve moves through your interface and does it on screen, at human speed.
“how do I invite a teammate?”
Show →03
↑ climbs when it needs your backend
Do
Peeve executes the action, after confirming the exact effect first.
“cancel my subscription.”
Do →04
↑ climbs when it can't resolve safely
Escalate
Hand off, in-product
Peeve opens a ticket carrying the full session, rather than guess.
“I was charged twice.”
Hand off →Peeve never climbs past what's safe. Unsure between showing and doing, it shows. It won't execute something you asked it to explain.
What you're replacing
One agent instead of five tools that don't talk to each other.
Deflection
A chat widget that describes the answer
Peeve shows it on the live screen, and only bills when it resolves.
Onboarding
An authored tour that breaks every deploy
Peeve walks users through your real product, generated live, never re-recorded.
Analytics
A separate stack to guess where users stall
Every session already measures the routes they get stuck on, in their words.
Escalation
Routing rules that hand over “user needs help”
Hand-offs carry the full session to your team in Peeve Hand-Offs, and into your helpdesk if you keep one.
Lead capture
A sales chatbot bolted onto your marketing site
Peeve catches buying intent inside the product and qualifies it in the moment.
Five subscriptions, five dashboards, five things to keep current, or one install where a single registry feeds every surface and one loop keeps it true.
Does real backend work
It doesn't just answer. It works your backend.
Connect your own tools two ways (your MCP server, or any REST API) and mid-conversation Peeve reads live data and takes real actions for your customers. No SDK, no integration project: minutes, not weeks.
Custom API ConnectorScale
Paste your API docs and Peeve builds the connector itself: base URL, auth, and every endpoint, each marked read or write. Reads run live to answer with current data; writes, like a refund, always run through a confirmation step and a workspace allow-list.
Custom MCP ServerGrowth and up
Already run an MCP server? Peeve plugs straight in. A live handshake discovers the tools it exposes, and you choose which the assistant may use. Anything that changes state runs through the same confirm guardrail.
You pick what's enabled
Turn individual tools and endpoints on or off; the auto-built connector is fully reviewable.
Credentials stay yours
Encrypted per workspace and never shown again. The model never sees a token.
Two kinds of users
Your users have agents now. Peeve serves those too.
Humans
The cursor moves, narrates, and pauses before it acts. Rendered as motion, in the session they're already in.
Agents
POST /mcp/tools/call
{ "tool": "invite_teammate",
"input": { "email": "sam@northwind.io",
"role": "editor" },
"dry_run": true,
"effect": "Adds 1 seat. Bills $12 on Nov 1." }
The same capability, rendered as a typed tool with its effect stated before it runs. Agents get the same permissions as your users, never more.
One capability registry. The cursor renders it as motion. The MCP server renders it as tools. Neither is maintained separately.
The agent channel →Agent integrations
Bring your own agent. Peeve serves it the same answers.
Claude, OpenAI, Hermes, OpenClaw (or any MCP client) connects to a scoped endpoint and gets your capabilities as typed tools, each stating its effect before it runs.
Claude
Add Peeve as a custom MCP connector.
OpenAI
Connect over MCP; call your tools.
Hermes
Connect the Hermes agent over MCP.
OpenClaw
Connect the OpenClaw agent over MCP.
Self-healing
Rename a button. Peeve notices before your users do.
Every AI support tool needs knowledge, and knowledge rots. Peeve treats knowledge freshness as the product, not maintenance work left to you.
Every session scores how confidently Peeve found each element. When a score falls, that route is flagged, re-captured on the next visit, and diffed against the map. A rename merges automatically. Anything structural waits 30 seconds of your attention.
Chat widgets can't do this. They never touch your interface, so they have no signal to collect.
How the Brain works →verified
flagged
not mapped
Peeve, by the numbers
~150ms
to route each question: answer, show, do, or escalate
$0.06/conv
per resolved conversation on Growth, against Fin's $0.99
0.60
confidence floor: below it Peeve hands off, never guesses
1 day
to install: one command, then a 20-minute walkthrough
Intent detection
Every pricing view and cancel click, tied to real revenue.
Peeve reads intent straight from your existing pricing cards and cancel buttons. No events to wire up. You see at-risk and upgrade-ready users the moment they act, even if they never open the assistant.
your pricing card: markup you already have
Growth
$99/yr
data-peeve-plan="Growth"
data-peeve-amount="$99.00"
data-peeve-currency="USD"
data-peeve-interval="yearly"
data-peeve-amount="$99.00"
data-peeve-currency="USD"
data-peeve-interval="yearly"
passive signal
→the signal Peeve captures
{
kind: "pricing_view",
plan: "Growth",
amount: 99,
currency: "USD",
interval: "yearly"
}
kind: "pricing_view",
plan: "Growth",
amount: 99,
currency: "USD",
interval: "yearly"
}
Zero-code capture
Add a few data-peeve-* attributes to markup you already have. No SDK calls, no event tracking.
Passive by design
Signals fire on a pricing view or a cancel click, even with the assistant closed.
Revenue-aware
Every signal carries the plan, amount, currency and interval, so intent ties to the money at stake.
Goal-aware conversations
Peeve knows whether each chat actually solved the user's problem, or was abandoned.
Try it on your own product
Not ready to install? Start with a number.
Two free tools, no signup: see how stale your help center is, and what your support questions cost today.
Install
One command. No migration project.
Then a 20-minute guided walkthrough where you use your own product once while Peeve records the routes.
Read the developer docs →$ npx peeve init
# detects your framework, adds the snippet, writes peeve.config.json
$ npx peeve verify
# 48 routes · good accessibility data · 2 canvas regions flagged
Loading and status lines say what they found. Never just that they're working.
Peeve acts in your product. It never guesses.
Every capability carries a confidence score; below 0.60 Peeve stops and hands off rather than guessing. Every backend action is confirmed before it runs, and Peeve stays read-only until you widen it.
Pricing
Published, because everyone else hides it.
14 days free, no card. Annual is two months free. Hard cap by default. You're never billed for overage you didn't buy.See the full pricing page →
Every planAPI, MCP and webhooks on every plan, including the free trial. Fair-use rate limits apply.
Security & compliance
Enterprise-grade from day one, and honest about what's in flight.
Every status carries a word, never a colour alone. Nothing is claimed before it is true.
SOC 2 Type IIAttestationIn progress · Q1 2027
GDPRCompliantSub-processor list published
Signed DPAData processingAvailable on Scale
EncryptedData protectionAt rest + in transit
Workspace-isolatedTenancyNever trains outside models
Encrypted at rest + in transitWorkspace-isolated, never trains outside modelsRead-only until you widen itKill switch in one commandSOC 2 in progressGDPR-ready
HIPAA / BAA is not offered today. The full sub-processor table, data map and retention defaults live in the trust center and on the security page.
Questions people ask.
Each answer is 40–60 self-contained words: subject named, no pronouns pointing backwards, one concrete number.
What is Peeve?
Peeve is an AI customer support and experience agent that installs into your web product with one snippet. It answers questions, shows users how by moving through your real interface, does tasks on your backend with confirmation, and hands off to your team with the full session attached.
How is Peeve different from a chatbot?
A chatbot returns text describing where something is. Peeve moves a cursor to it and clicks. Both answer; only one of them acts inside the interface the user is already looking at.
What kind of app does Peeve work with?
Peeve is built for a SaaS product with an authenticated dashboard, frequent releases, and recurring how-to tickets. Inside the product it runs the full ladder: it answers, shows users how by moving the cursor, and does approved backend actions. On your marketing site and docs it answers and hands off. One brain powers both. It works on any web interface, so a marketing site or docs is fine, but not a pure native mobile app (no web interface to navigate) or a canvas-only interface. On the surfaces it supports, Peeve resolves support requests end to end and, when a request genuinely needs a person, hands off to your team in Peeve Hand-Offs with the full session, also syncing into the helpdesk you already run if you keep one.
How long does installation take?
An afternoon. One command (npx peeve init) detects your framework and adds the snippet, then a roughly 20-minute guided walkthrough records your routes while you use your own product once. Nothing is live until you turn it on.
Does Peeve need my users' passwords?
Never. The model never sees a password or a session token. It requests a capability by name and an execution service maps that to the call, so a prompt injection can ask for a capability but never for an arbitrary request.
What happens when Peeve doesn't know?
It stops and says so. Every capability carries a confidence score; below 0.60 Peeve hands off rather than guessing, and conversations where it found nothing and said so are never billed.
Can AI agents use Peeve?
Yes. Your users connect their own Claude, OpenAI, Hermes, OpenClaw or custom agent to a scoped MCP endpoint and get your capabilities as typed tools with the effect stated before each runs. Agents get the same permissions as your users, never more.
What does Peeve cost?
Three public plans: Starter $40, Growth $199 and Scale $499 a month. One pool of credits powers every channel (a conversation is 1 credit per 24-hour window, voice is 0.5 credit a minute) with a hard cap. At typical completion that works out to about $0.06 per resolved conversation on Growth, against Intercom Fin's $0.99 per resolution.
Does Peeve work with Intercom or Zendesk?
Yes. Peeve imports your Intercom or Zendesk help content so it starts from what you have written, then resolves support requests end to end inside your product. When a request genuinely needs a person, Peeve hands off to your team in Peeve Hand-Offs and also syncs a Zendesk ticket with the full session, opening a Linear issue for bugs, so your existing tools keep working and their bill goes down.
Why we built this
I ran support on Intercom with around 150 help articles. The product kept changing; the articles did not. So every AI support tool leaned on stale knowledge, and users asking a simple how-to question got a long paragraph describing an interface that had already moved. Reviewing those conversations, finding what failed, and rewriting the help center became its own full-time job.
The answer to a how-to question should not be a wall of text. So we stopped describing the interface and started reading it. Peeve builds a living brain of the product, then answers, shows, does, or escalates, depending on what is safest and most useful. When it cannot do something safely, it says so and hands off. It never guesses.
Arslan Nasir
Founder, Peeve
14-day trial, no cardInstalls in an afternoonRead-only until you widen itKill switch in one commandCancel anytime. No ONE question asked.
The one question we'll ask on the way out: which answer did Peeve get wrong? It is how the Brain learns. And you were never billed for it in the first place.