Pushary for agent builders
Let your users approve what your AI agent does.
Your agent asks before a refund, an email or a delete. Your user taps Approve or Deny on their phone, under your brand.
What your user sees. Acme is your product.
TypeScript · Python · Vercel AI SDK · LangGraph · Mastra · OpenAI Agents SDK · CrewAI
Your framework can pause your agent. It can't reach your user.
Reaching them takes all of this. Pushary already has it:
- A phone app with Approve and Deny on the lock screen
- Push notifications for iPhone, Android and browsers
- A way for your users to connect without another account
- A request that stays open when your server restarts or deploys
- A timeout that means no, never yes
- A record of who said yes, and when
You add two calls to your code. That is the whole build.
Your agent asks. Your user decides.
Four steps, from connecting a user to running the action.
01
Connect your user once
Show them a link in your product. They install the free Pushary app, open the link and allow notifications. No sign-up, no payment.
02
Your agent asks first
Right before the risky step, your code sends the question to that user: “Approve a $4,200 refund?” It can be a yes or no, a choice from a list or a typed answer.
03
Your user taps a button
The notification starts with your product's name. They tap Approve or Deny on the notification.
04
Your agent gets the answer
Approve, and the agent carries on. Deny or no answer, and it stops. Every answer is saved, and can be sent to your server.
No sign-up. Takes a minute.
Your brand on every request.
Your users know your product, not ours. So every request leads with your name.
Acme needs your approvalnowApprove a $4,200 refund for order #4471?Shows your brand
- The notification
- Starts with your product's name: “Acme needs your approval”.
- The approval screen
- Your name and logo above the question.
- The web page
- Each request has one, with your name and logo. Switch off “Powered by Pushary” to drop our mark.
- The connect link
- Your name and logo, and a button to get the app.
Shows Pushary
- The app
- Your users install the free Pushary app once. Its icon shows on the notification, and some phones show its name too.
- The web address
- Links and web pages stay on pushary.com.
Your users never make a Pushary account and never pay us. One free app works for every product on Pushary, so you never publish an app of your own.
Only ask when it matters.
Set a few rules so your users are not asked about every small thing. Change a limit without shipping code.
- $42 refund
- Runs on its own. Nobody is asked.
- $4,200 refund
- Asks your user first.
- $40,000 refund
- Blocked. It never runs.
refund.create amount < 100 allow
refund.create amount >= 1000 require_approval
refund.create amount >= 10000 denyThis is shorthand. You add each rule with one API call. No rule for an action? It asks a person. A missing rule is never a yes.
Two calls. That's the integration.
Connect a user once, then ask before the risky step. TypeScript or Python, with adapters for the agent frameworks you already use.
Listed in the Vercel AI SDK tools registry, LangChain's integrations, the Hermes Agent plugin catalog and the Strands Agents integrations.
import { createPusharyServer } from "@pushary/server"
const px = createPusharyServer({ apiKey: process.env.PUSHARY_API_KEY! })
// 1. Once per user: show them this link in your product.
const { universalLink } = await px.enroll(user.id)
// 2. Before the risky step: ask, and wait for the answer.
const decision = await px.authorize({
toolName: "refund.create",
toolTarget: "order_4471",
externalId: user.id,
parameters: { amount: 4200 },
question: "Approve a $4,200 refund for order #4471?",
})
if (!decision.approved) throw new Error(decision.reason)
await issueRefund()- approved: true
- Your user tapped Approve, or a rule allowed it. Run the action.
- approved: false
- Your user said no, a rule blocked it, or nobody answered in time. Don't run it.
Who it's for
You ship an AI agent to customers
Your agent issues refunds, sends emails or changes accounts. Your customers approve the risky ones from their phone.
You build agents for clients
Stop rebuilding the approval step on every project. Add Pushary once and reuse it.
You run a platform others build on
Give every builder on your platform phone approvals, without building them yourself.
One plan covers one product. Building for several clients, or running a platform? Email us and we will set up a plan that fits.
Partner plans
Pick the size that fits your product. Existing Partner customers keep their terms.
Your first trial covers 100 of your users and 100 approval requests for 3 days. Paid limits start when billing starts. Retries and extra devices don't count as extra requests. Plain notifications without a question aren't included. No automatic overage charges.
Questions, answered
How do I let my users approve what my AI agent does?
Do my users have to sign up or pay?
Is it white-label?
Who gets asked?
Won't my users get asked all the time?
What if my user never answers?
How much is the Partner plan?
Which frameworks does it work with?
How is this different from Knock or Novu?
How is this different from HumanLayer?
I build agents for several clients. Do I need one plan each?
Does this cover the EU AI Act human oversight rule?
Your agent asks. Your users decide.
Two calls in your code. Your users answer on their phone, under your brand. Every answer is saved.
Start freeBy product type
Using AI agents yourself? Get their questions on your phone or your Mac, and answer from either one. See the agent plans.