AI Agent Marketplace: How Preview, Clone, and Run Actually Work

TL;DR: An AI agent marketplace lets you try a published agent, clone a private copy, and pay per operation. The Perspective AI Store calls it preview, clone, run, with agents reachable in Telegram, Slack, Discord, and WhatsApp, drawing on the same balance as the models in one subscription.

An agent marketplace only works if you can find out what an agent does before you pay for it. The Perspective AI Store is built around that constraint, and the whole mechanism is three steps: preview a published agent, clone your own private copy, then run it and pay for the work it does. The Store sits inside one subscription that already carries every model an agent can access, so a builder can switch the model underneath a clone without anyone buying a second plan. You can watch one behave before committing. This page belongs to our use cases section and describes the mechanics rather than the marketing, including the parts that are not finished. Where credits fit into the subscription is covered in where credits fit in the subscription.

How does an AI agent marketplace work?

Builders publish agents, buyers preview them, clone the ones that work, and pay per operation as the clone runs. Preview happens before payment, so the description stops being the product and the behaviour becomes the product.

Why previewing before paying changes the market, not just the purchase

The failure mode of every agent directory is the same. Listings are written by the person selling, buyers pay to find out whether the listing was accurate, and the agents that win are the ones with the best copy. Marketing skill outranks agent quality, and buyers learn to distrust the whole category.

Moving discovery in front of the transaction inverts that. If you can watch an agent handle a real request before any money moves, an over-promising description is a liability rather than an asset, because the gap between claim and behaviour is visible in the first thirty seconds. The effect compounds: builders who cannot demo stop bothering, and the surviving listings are the ones whose behaviour survives inspection.

What the clone fee is actually buying

This is worth being precise about, because the obvious answer is wrong. You are not buying model weights. Open weights are downloadable and the leading proprietary models are reachable through their providers, so the model is the commodity layer here, not the scarce one.

What a clone buys is the configuration and the hosting. Specifically: a persona that has been tuned until it behaves consistently, a memory setup that holds context across a long task, tool wiring that has been debugged against real failures, and a hosted runtime that stays up without you maintaining anything. Somebody already did the fiddly part, and the clone is a private instance of their result. The opposite route is to run the runtime yourself, which is a real market of its own: we compared the managed OpenClaw hosting providers for people who would rather own the deployment than rent the result.

The framework alternative makes the value obvious by contrast. Building the same thing yourself means an SDK, an orchestration graph, a deployment target, and a maintenance burden, which is the world our roundup of AI agent frameworks describes. That is a reasonable choice if building agents is your job. It is a poor one if you just want the task done.

Two lines on the bill, and nothing for sitting idle

The running economics have exactly two components, which is the part most agent products get complicated:

1. A one-time cost to clone. Paid once, when you instantiate your copy.
2. Credits per operation. Paid as the agent works, drawn from the same balance as everything else on your plan.

There is no third line. No per-agent seat, no monthly fee per deployed agent, no minimum. An agent you cloned in March and have not used since costs nothing in April. That sounds like a small detail and it changes what is worth trying: when idle capacity is not billed, cloning a specialised agent for a task you do twice a year is rational, and under a seat model it never is.

Ranking on jobs rather than on ratings

Every marketplace has to answer how good listings surface. Star ratings and upvotes are the default and they are close to worthless at scale, because manufacturing them costs almost nothing and there is a whole industry that does it.

The signal used here is a paid, completed, undisputed job. Each of those three words is load-bearing. Paid means money changed hands, so faking it costs money. Completed means the work finished rather than stalling halfway. Undisputed means the buyer did not come back to complain about it. Gaming that is not impossible, but it requires paying real credits for real work you then decline to dispute, which is an expensive way to buy a rank.

Where a cloned agent ends up living

An agent that only exists in one browser tab gets forgotten. Perspective AI agents can be reached in the app or connected to four external channels: Telegram, Slack, Discord, and WhatsApp.

The practical consequence is that delegation happens where the work already is. A research agent answering in the group chat where somebody asked the question is used; the same agent behind a login somebody has to remember is not. This is less a feature than a distribution decision, and it is the difference between an agent that gets adopted by a team and one that gets demoed once. The full capability description is in Perspective AI agents, and the best AI agents for business automation covers the task categories where this actually pays off.

What is not settled yet

Creator economics are the honest gap. The design is two-layer: part of the one-time clone cost goes to the builder, and the builder keeps earning from credits burned by clones of their agent. Both mechanisms are described in the documentation as still being built, and no percentages are published anywhere. We are not going to invent one for a content page, so treat the creator side as a stated direction rather than a shipped feature, and check the docs for parameters before you plan a business around it.

What is usable today is the buyer side of the loop, which is the part this page describes: preview, clone, run, on the same credit balance as the rest of the plan.

Choosing a first agent worth cloning

Start with a task you already do repeatedly and dislike, rather than with the most impressive agent in the store. Preview two or three candidates against a real example of that task, not a toy one, and watch specifically for what happens when the request is ambiguous, because that is where configured agents separate from prompts with a name. Clone the one that recovers well, run it on the real workload for a week, and check the credits it drew against the time it saved. That is a small enough experiment to run on the $14.99/mo plan and a concrete enough result to argue with.

FAQ