Claude Marketplace: Buying Agent Tools From Claude Spend

Anthropic launched Claude Marketplace on September 23, 2026. It is a single catalog where teams find connectors, agent products and service partners that work with Claude, and where builders and partners reach Claude customers. The line that matters to a buyer is one sentence in the announcement: teams can use a portion of their committed Anthropic spend on marketplace purchases.

That sentence moves an agent purchase from a new budget request to a drawdown on money already approved. It is the same mechanism the cloud marketplaces have used for years, and it is why a marketplace listing is a procurement event, not a directory entry. The rest of this post is about what changes for the buyer and what to check before drawing down.

Everything here comes from Anthropic's announcement. The percentage of committed spend that can be applied is not stated, and we do not guess at it.

The announcement describes the marketplace as live on the day. For builders the route in is to create a connector or plugin with MCP and Agent Skills and submit it, to apply to list a Claude-powered product, or to join the partner network for a services listing. Our earlier post on agent marketplaces as distribution covers how these catalogs have worked at other vendors; the new element here is the money.

A committed-spend contract is money a company has already agreed to pay a vendor over a period, usually in exchange for a discount. Once signed, it sits on the finance side as a line that will be spent. Letting a customer apply part of that line to third-party purchases changes three things at once.

First, the approval is already done. A team buying an agent product through the marketplace is not asking for new money; it is choosing how to use money the company has committed to Anthropic. The request that used to need a business case now needs a click and a check that the drawdown is within the allowed portion. Second, the contract path shortens. Purchases that would each have gone through vendor onboarding, security review and a separate invoice arrive through one relationship. Third, the incentive on the buying team flips. Unused commitment is a cost; a marketplace turns it into something the team can spend before it expires.

All three make it easier to buy. None makes the purchase safer, which is why the next section exists.

A marketplace listing built on MCP is an MCP server with a storefront. Every question we already ask of a connector applies. The routing below adds the ones that the money introduces.

Two of our reference pages carry the detail. The MCP security incident ledger records what has gone wrong with connectors in production, and the vendor claim due-diligence checklist is the list to run on any product before the drawdown.

The third category is the unusual one. Consultancies do not usually sit in a software catalog next to the software. Putting them there does two things for a buyer. It makes an implementation partner a line item in the same purchase as the product, which suits companies that want one decision rather than two. And it invites comparison between a services engagement and a product on the same page, which favors firms that can describe an outcome rather than a headcount.

For the service firms themselves, it sets the terms of entry: the partner network, not a direct listing. Anthropic names three large consultancies at launch. Whether smaller firms appear, and on what terms, is not stated in the announcement, and this post makes no claim about who will or will not be listed. What is clear is that the catalog rewards firms whose work can be bought in units, which is a shift many agencies, including ours, are already making in how we package AI transformation work.

A committed-spend line sits with finance. The product it is about to pay for sits with whoever will use it. The connector it installs sits with security. The contract it creates sits with legal. The marketplace puts all four behind one button, and the question a company has to answer before the first purchase is which of them gets to press it, and on whose sign-off.

The practical answer is that the budget owner authorizes the drawdown and the usual reviewers authorize the purchase, in that order. A team that wants to spend unused commitment proposes a listing; security runs the connector checks in the routing above; legal confirms who the data-processing agreement is with; the budget owner confirms the portion is within what the account team put in writing; then the button is pressed. Written down, this is the same process the company already runs for any software purchase. The marketplace only makes it faster to skip.

Decide the two edge cases in advance. First, whether unused commitment near a renewal date may be spent on the marketplace at all, or whether that is a case-by-case call. Second, who can veto: a purchase that passes finance and fails security should stop, and the process should say so before the first time it happens.

None of this requires a committee. For a small company it is one person wearing two hats and a checklist they agree to run before clicking; the point is that the checklist exists and that the budget line does not shortcut it. The marketplace will make skipping it feel efficient, because the money is already there and the catalog is one click away. The checks are the part of the purchase that now feels harder to justify, and that is the signal to protect them.

Each illustrative case below is a team using the marketplace exactly as designed and ending up somewhere it did not choose. None requires anything to be wrong with the catalog; each is a consequence of the budget line making a decision that another function should have made.

The budget-led buy: commitment is expiring, a product looks useful, the drawdown clears, and the data-processing terms with the product vendor are read for the first time at renewal. The assumed umbrella: a team takes Anthropic’s own terms as covering a third-party product bought through Anthropic’s catalog, when the listing says the contract is with the vendor. The unchecked reach: a connector is installed because it was in the catalog, and nobody asked which systems it can read or write once it is running inside Claude. The expiring access: a product paid for by drawdown stops when the commitment ends because the listing said nothing otherwise, the renewal is negotiated on other grounds, and the team discovers the product was tied to the contract rather than bought outright.

All four are prevented by the routing in the checks section, run before the drawdown rather than after. The marketplace shortens the path to buying. It does not shorten the list of things worth knowing before you do.

A useful habit is to write the one-paragraph answer to each of the four routing questions into the purchase record before the drawdown, so that the failure mode is at least documented if it happens. A record that says “contract is with the vendor, data stays in their region, connector reads two systems and writes none, drawdown is within the written portion” is a purchase someone can defend six months later. A record that says “bought via marketplace” is not.