About Pushary

AI agents act. Pushary decides what they may do.

Pushary sits between an agent and the real world. Risky actions get checked against your rules, and when a rule cannot decide on its own, a real person does, on their phone, in seconds.

What Pushary is

Pushary is the permission layer for AI agents. Before an agent runs a command, spends money, sends a message, or ships code, Pushary checks your rules and gives one of three answers: allow, deny, or ask a human. Asking a human means a push notification on a real phone, an answer in one tap, and a permanent record of who decided what.

New to the idea? Human-in-the-loop for AI agents, explained covers what it means and where the checkpoints belong.

Why we build this

Agents used to suggest. Now they act.

An agent that can only write text is safe by accident. An agent with a shell, a card on file, and your production keys is a different thing entirely.

The blast radius is real now

Agents run shell commands, move money, email your customers, and deploy to production. A wrong step is no longer a bad paragraph you can delete. It is a dropped table, or a refund that should never have gone out.

Both of the usual answers are bad

You can watch every step, which turns a fast agent into a slow version of you. Or you can let it run and hope, which works right up until the day it does not. Neither one survives past one person at one keyboard.

Afterwards, nobody can say what happened

Chat history is not an audit trail. When someone asks who approved the refund, or why the agent was allowed near that server, there is usually no answer to give.

The policy engine

One rule, three possible answers

A policy is a rule you write once. Reading files is always fine. Deleting things always asks. A refund over five hundred goes to a manager. Pushary reads that rule every time an agent wants to act, before anything reaches the outside world.

Ask a human

The rule hands it to a person

The question goes to your phone and the agent waits. Approve or deny confirmation questions from the notification. Open the app to choose an option or type a reply. Your answer goes back to the agent.

You decide what silence means

Every rule carries a timer and an instruction for when it runs out: keep waiting, deny, pass it to someone else, or allow. Silence never becomes an accidental yes unless you wrote a rule saying it should.

Allow

The rule already says yes

Nobody is interrupted and no notification is sent. The agent keeps moving. The allow is still written to the log, so the record stays complete even when no human was involved.
Deny

The rule says no

The agent is stopped and told why, in words a model can read, so it takes a different route instead of trying the same thing again.

One set of rules, every agent

The same policy covers Claude Code, Codex, Gemini CLI, and Cursor, plus the SDKs your own product runs on. You write a rule once instead of learning four different permission screens.

Claude Code, Codex, Gemini CLI, Cursor, Hermes, and the Claude and ChatGPT connectors. For your own product: a TypeScript SDK, a Python package, a REST API, and a hosted MCP server.

What is shipped, and what we are still building

The phone approval is the part people adopt in five minutes. The policy engine is the part that makes it hold up at a hundred agents and a thousand users.

Shipped

On your machine, for the agents you run

A small hook installs next to your coding agent and evaluates your rules locally. An allow costs nothing and needs no network. Anything that needs a person reaches your phone, and every outcome lands in a log that cannot be edited afterwards.
Being built

On our servers, for the products you build

The same engine is moving server side, so any product can ask for a verdict over the API and get allow, deny, or ask a human for its own end users. Same rules, same record, no laptop in the path.

Every decision is written down, and stays written down

The log is append-only, enforced by the database itself, so a row cannot be edited or quietly removed later.

  • What was asked, and which action it was guarding
  • Who answered, or which rule answered instead of a person
  • How long the agent waited for it
  • How it ended: approved, denied, timed out, or answered

Two different people buy this

One is a developer who wants their own agents to reach them. The other is a company whose agents need to reach their customers.

You, watching the agents you run

Your agents run on your machine and you want them to reach you instead of stalling or guessing. You approve from your phone, the rules follow you between agents, and the history is yours.

Agent

For one person, every agent

$9.99per month

  • 5K notifications/month
  • Every agent you run
  • Approve, deny, or answer from your phone
  • Mac app included
  • Permission policies
  • Full MCP and API access
  • 30-day history

Agent Pro

For power users

$19.99per month

  • Unlimited notifications
  • Budgets
  • Priority support
  • For one person. Team adds people and routing.
  • Policy Autopilot suggestions
  • 90-day history
  • Everything in Agent

Team

For teams that need a shared record

$59per month

  • Everything in Agent Pro
  • 5 people included; add more in Billing
  • Shared audit trail with export
  • Team activity and a summary email
  • 365-day history
  • Invite your security team

Your product, asking your own users

You are building an AI product and the person who should approve is your customer, not you. Pushary runs inside your product. Your users connect a phone with one tap, never see a Pushary login, and never get a bill from us.

Partner Launch

Approvals inside your product

$99per month

  • Your users approve what your agent does
  • Up to 5K of your users
  • 25K approval requests a month
  • 3 operator seats
  • 90 days of answer history
  • Every answer sent to your server
  • TypeScript, Python, REST and MCP

Partner Growth

Approvals inside your product

$299per month

  • Your users approve what your agent does
  • Up to 25K of your users
  • 100K approval requests a month
  • 10 operator seats
  • 365 days of answer history
  • Every answer sent to your server
  • TypeScript, Python, REST and MCP
  • Priority support

Partner Scale

Approvals inside your product

$799per month

  • Your users approve what your agent does
  • Up to 100K of your users
  • 500K approval requests a month
  • 25 operator seats
  • 730 days of answer history
  • Every answer sent to your server
  • TypeScript, Python, REST and MCP
  • Priority support

Enterprise

Custom agreement

Customlet's talk

  • Your users approve what your agent does
  • Every answer sent to your server
  • TypeScript, Python, REST and MCP
Compare every plan

Every agent plan starts with a 3-day trial.

Who is behind it

Pushary is built and operated by RalphNex OÜ, registered in Tallinn, Estonia under registry code 16932562 and founded in 2024.

It is a small team led by Aadil Ghani, who founded the company and still writes the code, which is why the docs read like a README and the changelog moves every week.

Where your data lives

Pushary is operated by an EU company. Processing locations depend on the services involved; EU incorporation does not mean all data stays in the EU. Contact us to review your requirements. We do not claim a certification we have not obtained.

What we can and cannot see

API keys are stored only as hashes, so we cannot read yours back to you. Approval requests carry the question, your answer and the command or change you are asked to approve, with known secrets removed. Session transcripts are redacted and encrypted on your machine; your phone can decrypt them, and authorized Pushary compliance admins can recover them for documented legal obligations.

How workspaces are kept apart

Every database query is scoped to the workspace that owns the data, so one customer's rules and decisions are unreachable from another's.

Talk to us

Questions about the policy engine, the Partner plan, or something that went wrong go to a person, not a queue.

Business and partnerships
business@pushary.com

See what your agent is about to do, before it does it.

Connect an agent, write one rule, and take the next approval on your phone.

Get started

Agent plans from $9.99 a month. Partner plan for embedding approvals in your own product, $99 a month.