Strip away the vocabulary and a sales agent is a web server you run at your own domain. It publishes a manifest at a well-known URL so buyer agents can find it, exposes your inventory as a set of callable tools rather than a rate card, and answers a question like "what video inventory do you have in the UK next month and what does it cost" with a structured reply another machine can act on. The Prebid Sales Agent is the open-source reference version of that server, maintained by the organisation behind header bidding, and a publisher with Docker installed can have one running locally in minutes. Whether that is worth doing yet is a separate question, and the honest answer depends less on the software than on whether any buyer is calling.
What is the Prebid Sales Agent?
The Prebid Sales Agent is a server that exposes advertising inventory to AI agents through the Model Context Protocol (MCP) and the Agent-to-Agent (A2A) protocol, integrates with ad servers such as Google Ad Manager, and provides tools for managing inventory and campaigns across their lifecycle. That description is Prebid's own, from the project documentation at docs.prebid.org.
Prebid.org announced it on 29 January 2026, in collaboration with AgenticAdvertising.org. The code had been incubated inside the agentic advertising community and was built on the Ad Context Protocol, the open standard launched on 15 October 2025 by a consortium including Scope3, Triton Digital, Yahoo, PubMatic, Optable and Swivel. Prebid took stewardship of the repository, open-sourced it and released it to any publisher who wants it, with no membership required. Mike Racic, then president of Prebid.org, said the collaboration "makes agentic advertising practical for publishers of any size".
The governance split matters and is deliberate. Prebid leads software development and maintains the Sales Agent repository. AgenticAdvertising.org governs the AdCP specification and compliance. Garrett McGrath, then chairman of the Prebid board, put the distinction to AdExchanger plainly: "We're not a standards org; we're a software org."
It is a reference implementation, not a product. McGrath was explicit that it would not be one-size-fits-all, and that publishers would adapt it to their own needs in the same way they adapted Prebid's header bidding software. Building agents from a shared reference is meant to stop the ecosystem repeating the early header bidding problem, where every publisher needed a bespoke integration with every demand partner.
How does the Prebid Sales Agent actually work, step by step?
This is the part the market keeps skipping, and it is the reason adoption is slower than the announcements suggest. Nikita Bansal, chief executive and marketing officer at TeqBlaze, argued in The Current in March 2026 that publishers are not stuck on which standard should win. They are stuck because most of the market cannot describe the interaction flow: who initiates a request, what is inside it, who responds, and where verification, approval and execution happen. "This is not a protocol-specific problem; it is a market-level transition issue," Bansal wrote. "What is missing is not the protocol. What is missing is translation."
So, concretely. The agent is written in Python and serves two interfaces from a single deployment.
The MCP interface presents your inventory as structured tools and resources that an MCP-capable AI assistant can call directly. Prebid's documentation names Claude Desktop as the example of that integration path.
The A2A interface is a JSON-RPC 2.0 server for autonomous agent-to-agent workflows that do not route through a specific assistant. Discovery happens through a standard agent manifest published at a well-known path, so buyer agents scanning for seller endpoints find yours without an introduction.
On top of those transports, AdCP defines the advertising operations themselves. Prebid's documentation describes four core domains: inventory discovery via get_products, where an agent searches in natural language rather than by line item ID; media buying via create_media_buy, a normalised proposal, negotiation and booking workflow that behaves the same across different ad servers; creative management via build_creative; and signal activation via get_signals and activate_signal for passing context and identity signals into targeting.
A campaign therefore runs roughly like this. A buyer agent discovers your endpoint and calls get_products with a brief. Your agent returns matching products. The buyer agent submits a proposal through create_media_buy. Your agent, and optionally you, review it. On approval, the agent pushes the order into your ad server. Delivery, pacing, reporting and reconciliation follow through the same protocol.
Two features are doing quiet work there. Approval is a first-class workflow state, so human-in-the-loop is a configuration choice rather than a bolt-on. And every exchange is a logged conversation between two named agents, which produces an audit trail neither side can later dispute. For publishers who have spent a decade unable to see who bought their open-auction inventory or why, that is arguably the more interesting property.
What can you sell through a sales agent that you cannot sell through OpenRTB?
The most useful answer is not "the same things, faster". It is the inventory that programmatic never learned to handle.
OpenRTB was designed for standardised units bought at the impression level. Custom sponsorships, newsletter placements, podcast reads, branded content and conversational or AI chat placements have stayed manual because there was no machine-readable way to describe them. AdCP negotiates in natural language over structured products, so anything a publisher can describe can in principle be transacted, and the specialist audio case is a live example: Benjamin Masse, chief product officer at Triton Digital, framed the sales agent as the route to automated buying for radio and podcasts without giving up the contextual depth that makes audio work.
This is also why the protocol is best understood as sitting above the auction rather than competing with it. Masse has characterised AdCP as "a protocol for investing" against the "day trading" of OpenRTB. Buyer agents engage your agent to structure larger pre-negotiated deals; the impression-level auction carries on underneath. In practice, agentic activity so far has clustered in direct and programmatic guaranteed deals. PubMatic has completed more than 1,000 agentic direct deals, and roughly half of the thirty or so agentic programmatic campaigns run through its platform used its Optable integration. McGrath expects Prebid's header bidding expertise to push agentic buying into the open auction eventually, but that has not happened yet.
How do you deploy one, and what does it actually cost?
Prebid publishes a Docker quick start: clone the repository, run docker compose, and the full stack including the database comes up on localhost, with the admin UI, the MCP server and the A2A server each on their own path. Cloud deployment guides cover the production case. BiddingStack, a commercial seller-agent host and an AdCP founding member, puts realistic timings on the three common paths in its publisher guide: roughly two minutes for local Docker, fifteen for Fly.io, twenty for Google Cloud Run. Treat those as a vendor's figures for standing up a default instance, not for going live.
Configuration is where the real work sits. You connect your ad server, with Google Ad Manager supported through the adapter pattern out of the box. You define products that map to your line items and inventory. You set pricing, targeting parameters and deal rules for each product. You test the endpoint against an AdCP-compatible buyer agent. Then you submit your discovery URL to buyer agent registries, because an endpoint nobody knows about is an endpoint nobody calls.
The software is free. The running cost is not. Self-hosting means you own uptime, scaling, security patching, ad server adapter compatibility, monitoring and incident response, and you own keeping pace with a protocol still under active development. That is the same trade publishers already make between self-hosting Prebid Server and paying someone to run it, and the same answer applies: if you have engineers and want control, self-host; if you do not, the managed hosting market exists and has an obvious commercial interest in telling you so.
Who governs this, and how stable is it?
Two caveats belong in any honest assessment, and neither is a reason not to look at the software.
The first is protocol competition. AdCP is not the only agentic standard. The IAB Tech Lab released its Agentic RTB Framework v1.0 for public comment in November 2025 and announced a broader agentic roadmap in January 2026, building on OpenRTB, AdCOM and VAST rather than starting fresh. Notably absent from AdCP's early adopter list were Google, Amazon, The Trade Desk and Microsoft, which between them operate the demand-side platforms handling most programmatic spend. Joel Meyer, OpenX chief technology officer and Prebid's chairman since August 2026, was refreshingly unsentimental about where this ends: "As an engineer, I will tell you it's always best if there is one protocol. I will also tell you, as a pragmatist, there will always be more than one." His stated position is that Prebid stays agnostic, and that its job is to "design software so they can sell their inventory regardless of the protocol being used to buy it".
The second is organisational churn. Racic, McGrath and Christian Janelli, the three people closest to the sales agent decision, all left Prebid in May 2026. Meyer told AdExchanger in August that the move was a good one but that the board had been "arm's length from it" and wants to get closer, and that Prebid hoped to name a new president in time for its summit in New York on 13 October 2026. A publisher betting operational workflow on an open-source project is entitled to ask who is maintaining it next year.
Should you run one?
The case for is that the cost of a small experiment is genuinely low and the cost of being undiscoverable is unbounded. Phil Bohn, senior vice president of demand at Freestar and Prebid's treasurer, made the adoption argument at launch: "Publishers gain a clear path to agentic advertising capabilities without needing to build from scratch or adopt vendor-specific solutions. It democratizes choice for publishers and accelerates ecosystem-wide adoption."
The case against is that demand is early and concentrated. Thirty campaigns through one large SSP is a pilot, not a channel. If you have no engineering capacity, no non-standard inventory that programmatic fails to sell, and no direct sales overhead worth automating, there is nothing wrong with deploying the agent in Docker, mapping two or three products, understanding the workflow, and waiting.
What is not defensible is treating this as a decision you can skip because the vocabulary is unfamiliar. Bansal's point stands: publishers who cannot describe the workflow end to end will not test confidently and will not push it internally, and that is a self-inflicted disadvantage rather than a technology risk.
Where a sales agent stops and the AI answer begins
A sales agent solves one transaction: an advertiser's machine wants to place an ad against your inventory, and your machine sells it the placement. That is a real and growing transaction, and having an endpoint that can service it is sensible infrastructure.
It does not address a different transaction that is now happening at scale on the same content. When an AI assistant retrieves your page to answer a user's question, no page renders, no ad slot exists and no bid request is made. There is no inventory for a buyer agent to buy, because the audience never arrives. That retrieval is where blankspace works: detecting live search agent requests at the CDN edge and placing contextual brand facts into the response the assistant is composing, which is a monetisation of the retrieval itself rather than of a placement on a page. The two are complementary rather than competing, and a publisher can reasonably run both. What neither can do is guarantee how a model uses what it is given.
Frequently asked questions
Is the Prebid Sales Agent free, and do I need to be a Prebid member?
Yes and no, in that order. The software is open source and freely available from the Prebid repository on GitHub. Prebid stated at launch that publishers can download it and contribute feedback regardless of whether they are Prebid members, and the repository is open for direct contributions without membership in either Prebid.org or AgenticAdvertising.org. Your costs are hosting, integration and maintenance, not licensing.
Does the Prebid Sales Agent replace header bidding or Prebid.js?
No. It is a separate channel operating above the auction layer, not a replacement for it. Prebid.js and Prebid Server handle impression-level auctions on your pages; the sales agent handles agent-to-agent negotiation of structured deals, which are then pushed into your ad server for delivery. Prebid maintains it alongside its existing codebases and supports it the same way.
Which ad servers does it support?
Google Ad Manager is supported out of the box through the project's adapter pattern, which is designed to accommodate additional ad servers. If your stack is not GAM, check the current adapter coverage in the repository before planning anything, because adapter support is one of the areas that moves fastest in an actively developed open-source project.
Can ChatGPT, Claude or Gemini buy my inventory through it?
Not quite as advertised. The agent exposes an MCP interface, and MCP is the standard those assistants use to call external tools, so an MCP-capable assistant can technically query your endpoint. Prebid's documentation names Claude Desktop as the example. That is a capability, not a demand channel: it does not mean those platforms' advertising businesses are buying media through your agent today. Vendor marketing tends to blur that line.
Does running a sales agent get me paid when an AI assistant summarises my article?
No. A sales agent sells advertising against inventory. If an assistant retrieves your content and answers the user without sending them to your page, there is no impression, no auction and nothing for a buyer agent to transact. Getting compensated for that retrieval is a different problem with a different set of answers, including licensing, pay-per-crawl and edge-level monetisation, and none of them is solved by having a sales agent endpoint.