AI Agent Marketplaces Compared: AWS, Google, Microsoft, Salesforce, and ServiceNow Agent Stores

An AI agent marketplace is a catalog of prebuilt AI agents you can browse, buy, and deploy without building anything. Think of it the way you think of an app store, with one important difference: an app runs on your device, while an agent needs access to your data and permission to act on it. That difference is what makes the choice of marketplace a bigger decision than it looks.

The category barely existed two years ago. It now includes catalogs from Amazon, Google, Microsoft, Salesforce, and ServiceNow, plus independent directories. They are usually described in the same language, so it helps to separate what they actually sell:

• Agent catalogs list complete agents that do a job end to end, such as qualifying inbound leads or reconciling an account. AWS, Google, and Agentmarketplace mostly sell these.
• Component catalogs list the parts an agent is assembled from: actions, topics, prompt templates, connectors. Salesforce AgentExchange is largely this, which is why its listings say "actions, topics, and templates" rather than "agents".
• Tool catalogs list the things agents call rather than the agents themselves. The AWS category is explicitly "AI Agents and Tools" for this reason, and much of it is MCP servers.

All three get filed under "agent marketplace" in press coverage. If you are evaluating one, the first question worth asking a vendor is which of the three you are looking at, because a catalog of 400 actions and a catalog of 40 finished agents are not comparable numbers even though both get quoted as marketplace size.

Here is the finding that matters most and that no vendor page will state plainly: every big-vendor AI agent marketplace is a distribution channel for its own platform, not a neutral market.

That is not a criticism, it is just what they are for. AWS built its catalog to pull workloads onto Bedrock. Google built its Agent Marketplace to make Gemini Enterprise seats stickier. Salesforce built AgentExchange to make Agentforce the place agentic work happens. Each one is doing the rational thing. But it has two consequences a buyer should price in.

The agent price is never the whole price. You cannot buy from these catalogs without first buying the platform. Google is the clearest example, because the gate is written into the edition table: Agent Marketplace access is a Plus feature, so the entry ticket is $50 per seat per month on an annual commitment or $60 without, and a company on the $30 Standard edition simply does not see the marketplace. Microsoft is similar in effect if not in structure. You need Microsoft 365 Copilot licences, and then agent usage draws on a separate Copilot Credit pool on top of those seats.

The configuration work does not travel. An agent bought from a vendor catalog runs in that vendor runtime. The month you spend tuning its instructions, connecting its data sources, and getting its approvals right is an investment in that platform, not in the agent. If you move, you rebuild.

None of this makes vendor marketplaces a bad buy. If you are already deep in Salesforce, AgentExchange is genuinely the cheapest and fastest place to get an agent live, because the expensive prerequisite is already paid for. The mistake is treating a vendor marketplace as a cheap way to try agents when you do not already run the platform. Priced honestly, it is a reason to buy the platform.

Yes. AWS launched an AI Agents and Tools category in AWS Marketplace on July 16, 2025, announced at the AWS Summit in New York. It is a centralized catalog where you search in natural language, review listings, and buy through your existing AWS account.

Two design decisions make it more open than most of its rivals. First, listings can declare support for Model Context Protocol (MCP) and agent-to-agent (A2A) standards, so what you buy is more likely to interoperate with things you did not buy from AWS. Second, the deployment targets are named and concrete: Amazon Bedrock AgentCore Runtime for the agent itself and AgentCore Gateway for adding tools, all running in your own AWS account rather than in a vendor-hosted app.

Commercially, you pay the seller price and AWS takes a transaction fee from the seller. Published rates are 3% on public SaaS offers and a tiered 1.5% to 3% on private offers depending on total contract value, with an extra 0.5% uplift on channel partner private offers. Those fees are charged on transactions, not on listing.

The catch is the obvious one. Everything assumes AWS. Buying is centralized through your AWS account, which is a real benefit for procurement (one bill, one set of controls, spend that can count toward committed AWS spend), and it is exactly the thing that makes it awkward if your data is not in AWS. For a company already committed to Bedrock, this is the strongest of the hyperscaler catalogs. For anyone else it is a large prerequisite.

Yes, and it is the newest of the group. Google announced Gemini Enterprise on October 9, 2025 as what Sundar Pichai called "the new front door for Google AI in your workplace". Partner-built agents from the Agent Marketplace became available inside the Agent Gallery on April 23, 2026.

The launch named twelve featured partners, including Accenture, Adobe, Atlassian, Deloitte, Oracle, Palo Alto Networks, Replit, S&P Global, Salesforce, ServiceNow, and Workday, alongside more than a hundred additional partner agents. Google also put a $750 million partner fund behind agentic development, which tells you how seriously it is treating the channel.

The governance model is the best-designed part. Employees browse the Agent Gallery and submit a request rather than installing whatever they like; IT approves or denies. Each agent gets a cryptographically distinct identity so there is an audit trail of what acted on what, and traffic passes through Agent Gateway and Model Armor. Google states that data is not used for model training. For a regulated buyer, that is a serious answer to the shadow-agent problem.

The pricing gate is the part to read carefully, because it is where most evaluations go wrong. Gemini Enterprise editions, as reported by Google Cloud resellers, run roughly like this:

• Business, about $21 per user per month with an annual commitment.
• Standard, $30 per user per month on a 12-month commitment or $35 without. Includes company-wide search, selected data connections, basic agent controls, and 30 GiB of storage per user.
• Plus, $50 per user per month with an annual commitment or $60 without. Adds the newest Gemini models, the Made by Google agents such as Deep Research and Code Assist, all data system connections, custom agent building in Agent Designer, 75 GiB per user, and Agent Marketplace access.
• Frontline, a lighter usage-capped tier for deskless staff that requires 150 Standard or Plus licences before it can be activated, and can only run pre-approved agents.

So the marketplace is a Plus feature. If you priced Gemini Enterprise at the widely quoted $30 seat and assumed partner agents came with it, your real number is $50 per seat per month on an annual commitment, plus compute consumption, which bills separately to your linked Google Cloud account and is not included in the seat fee. Google does not publish a public pricing page for Gemini Enterprise, so treat every figure here as reseller-reported list and confirm your own quote.

Yes. The Agent Store lives inside Microsoft 365 Copilot, where users browse, install, and try agents from Microsoft, from validated partners, and from their own organization, with the catalog surfacing across Teams and the Microsoft 365 Copilot app.

It has three distinct supply routes, and knowing which one an agent came from tells you a lot about how much it was checked:

• Independent software vendors submit an app package to Microsoft Partner Center through the Microsoft 365 and Copilot program. Microsoft validates it, the agent lands in the commercial marketplace, and it appears in the Agent Store.
• Copilot Studio makers publish an existing agent straight from the Channels section by choosing Microsoft 365 Copilot or Teams as the destination.
• Agent Builder users inside your own company submit agents to the "Built by your org" section, where an admin reviews and approves them in the Microsoft 365 or Teams admin center before they publish.

The admin approval step in that third route is the useful governance control, and it is the one most Microsoft shops discover late, usually after someone notices a dozen homemade agents in the tenant.

On cost, Microsoft stacks differently from Google. You need Microsoft 365 Copilot licences for the people using the agents, and then autonomous agent consumption draws on Copilot Credits, a separate tenant-wide pool. The rates are published in credits rather than dollars, which makes the real bill hard to see until you convert them, and we did that conversion in full in our breakdown of what Copilot Credits actually cost. The single most valuable rule there: employee-facing use by someone who already holds a Microsoft 365 Copilot licence is charged at zero, while customer-facing use bills every response. For seat pricing across the suite, see Microsoft 365 Copilot pricing.

These two are the enterprise application catalogs, and they behave more like each other than like the hyperscalers.

Salesforce AgentExchange launched on March 4, 2025 with more than 200 partners, including Google Cloud, Docusign, and Box. It is built on top of AppExchange, which Salesforce describes as the first and largest enterprise cloud marketplace with more than 13 million app installs, so the distribution machinery was already there. What you buy is mostly components rather than finished agents: prebuilt actions, topics, and templates that plug into an Agentforce agent you assemble, surfaced both on the marketplace and directly inside Salesforce agent-building tools. Everything requires Agentforce, and agent usage bills through Flex Credits or per conversation. We costed that model out in Agentforce pricing.

The ServiceNow Store added a dedicated AI agents category in January 2026 as part of a wider partner program overhaul. Partners build, test, certify, and distribute through the Build Partner Program, whose tiers run Registered, Select, Premier, and Elite, with a new Access tier that lets early-stage partners start building without full enrollment. The certification requirement is the point: this is a curated catalog aimed at enterprise IT and operations buyers, not an open directory.

ServiceNow also owns the fourth marketplace in this paragraph. Moveworks shipped an AI Agent Marketplace inside Agent Studio offering hundreds of prebuilt installable agents, and ServiceNow completed its $2.85 billion acquisition of Moveworks on December 15, 2025, the largest deal in its history. In February 2026 the two were combined into ServiceNow EmployeeWorks. If you are comparing Moveworks as an independent option, it is not one any more.

The practical read on both: if you already run the platform, these are excellent, because certification and native permissions are genuinely worth something and you have already paid the entry cost. Note that ServiceNow AI capabilities are not in the base licence, they sit in Pro Plus and Enterprise Plus, which we covered in AI agents for ServiceNow.

There is no single best one, because the honest answer depends entirely on a question about you rather than about them: what have you already bought?

Run through it in this order and the choice usually makes itself.

1. Is the process you want automated already inside one platform? If the work lives end to end in Salesforce, buy from AgentExchange. If it lives in ServiceNow, buy from the ServiceNow Store. Native agents see the platform permission model, inherit its audit trail, and cost you nothing extra in integration work. When the fit is genuine, nothing beats it.

2. Do you already hold the seats? A Microsoft shop with Copilot deployed can turn on employee-facing Agent Store agents at close to zero marginal cost, which is the best deal in this entire comparison. A company without those seats is looking at a per-user platform purchase before the first agent runs.

3. Does the work span systems? This is where vendor catalogs get weak, and it describes a lot of real business processes. Chasing an unpaid invoice touches a mailbox, an accounting system, and a bank portal. Screening candidates touches a job board, an ATS, and a calendar. No single vendor owns that chain, so an agent that only runs inside one of them can only do part of the job.

4. How sure are you? If the answer is "we want to find out whether this works before committing", the annual per-seat commitments attached to the vendor catalogs are the wrong shape for the question. That is the case for starting somewhere neutral and cheap.

That fourth situation is the one Agentmarketplace is built for. Agents deploy onto your own stack through scoped, revocable API permissions, run on whichever model you choose, and start at $29 a month with no platform seat underneath. You can run one agent against one process for a month and be wrong about it cheaply, which is a materially different bet from committing a year of seats to find the same thing out. If you want the buying checklist itself, we wrote it up in how to buy an AI agent, and the wider market pricing picture is on AI agent pricing.

This is the gap in the market, and it is worth being precise about what "neutral" has to mean for it to be real. Three tests separate a genuinely rail-neutral marketplace from one that just says so.

Where does the agent execute? If the answer is "in our runtime", you have bought a hosted product and the neutrality is cosmetic. A neutral agent connects to your systems and runs against them, so what you are buying is the logic and the connections, not a tenancy. On Agentmarketplace, agents deploy onto your own stack with scoped, revocable permissions, which means you can see exactly what an agent can touch and can withdraw it in one action.

Who picks the model? A marketplace tied to one model provider inherits that provider's pricing, availability, and roadmap. The GPT Store is the clearest illustration: an enormous catalog, trivially easy if you already use ChatGPT, and every item in it runs on OpenAI. Model-agnostic means you can move an agent onto a cheaper or better model when one arrives without rebuilding it.

What does leaving cost? Ask the question before you buy rather than after. If removing an agent means unpicking it from a platform you also use for something else, the switching cost is much higher than the subscription suggested.

Two practical notes for anyone shortlisting. Keep a human in the approval path for anything that moves money or emails a customer, at least for the first few weeks, regardless of which marketplace you buy from. And check the permission scope before deployment rather than after, since the most common failure in agent rollouts is not a bad answer, it is an agent with far broader access than the job required. Our own listings are organized by job to be done, so you can browse AI agents by category and see the permissions each one requests before anything is connected.