The Ad Context Protocol (AdCP) is an open technical standard that defines how software agents built on large language models buy and sell advertising. It specifies a common set of tasks, such as finding inventory, booking a campaign, uploading creatives and pulling delivery reports, so that a buyer's agent can talk to any seller's agent without a custom integration for each pair. It exists because every ad platform exposes a different interface, and an AI assistant asked to "find sports inventory for a running shoe launch" otherwise needs a bespoke connector for every publisher, supply-side platform (SSP) and data provider it touches.

What the protocol defines

AdCP does not move money or serve ads by itself. It is a vocabulary of named operations, described in JSON schemas, that agents call over existing transports. Planning-time work runs over the Model Context Protocol (MCP), the tool-calling standard Anthropic published in 2024, or over Google's Agent2Agent (A2A) protocol. The launch documentation said both transports carry identical functionality.

The specification is organised into protocol surfaces. As of October 2026 the project's repository lists:

  • Media Buy: inventory discovery, campaign creation and reporting (get_products, create_media_buy, get_media_buy_delivery).
  • Creative: building, previewing and syncing ads (build_creative, preview_creative).
  • Signals: audience and targeting data activation (get_signals, activate_signal).
  • Accounts: commercial identity and billing (sync_accounts, report_usage).
  • Governance, Brand and Sponsored Intelligence: suitability standards, a brand.json advertiser identity file and conversational ads inside AI assistants.
  • Trusted Match: a serve-time layer, discussed below.

At launch on October 15, 2025, the scope was narrower. PPC Land's report counted nine core tasks across discovery, comparison and activation, including list_authorized_properties, which checks which sites a seller may represent.

How a buy works, step by step

A typical transaction starts with a natural-language brief. The buyer agent sends get_products to a seller agent with the brief, the brand and a required buying_mode field. In version 3 that field takes one of three values: brief, where the publisher curates recommendations from the brief; wholesale, where the buyer asks for the raw product feed so it can apply its own audiences; and refine, which iterates on an earlier response.

The seller returns products, each with pricing options. The schema's ten pricing models run from CPM (cost per thousand impressions) and cost per click to cost per acquisition, revenue share and flat rate, for both fixed-price and auction buying.

The buyer then calls create_media_buy. The request must carry an account, a brand, start and end times and an idempotency key, a unique string that stops a retried request from booking the same campaign twice. Each package inside the buy names a product_id and a pricing_option_id, plus optional budget, pacing, bid price, frequency caps and creative assignments. Creatives follow through sync_creatives, and get_media_buy_delivery returns performance.

Behind the seller agent sits whatever system actually runs the campaign: the launch documentation mapped AdCP buys onto Google Ad Manager orders and line items, Kevel flights and Triton Digital campaigns. The protocol defines what must happen, not which ad server does it.

Tasks can return completed, working (expected within about 120 seconds) or submitted for jobs lasting hours or days, such as a human sales team reviewing a request.

Who operates it

On the sell side, a publisher, SSP or ad server runs a seller agent (AdCP calls it a sales agent) that answers requests against its inventory. Publishers are meant to declare which agents may sell their inventory in an adagents.json file, a mechanism Brian O'Kelley has pitched as a successor to ads.txt. The open-source reference sales agent moved to Prebid.org's stewardship on January 29, 2026.

On the buy side, agencies, demand-side platforms (DSPs) and vendors run buyer agents. Scope3, the company most associated with the protocol, acted as buyer agent in the first agent-to-agent test with Magnite's SpringServe seller agent in December 2025. Creative production can go to a separate creative agent.

Origin and evolution

O'Kelley, who founded AppNexus in 2007 and ran it until AT&T bought it in 2018, made the first commit to the project's public repository on July 19, 2025. The protocol launched as version 2.0.0 on October 15, 2025, with six founding members: Scope3, Yahoo, PubMatic, Swivel, Triton Digital and Optable. Magnite, Kargo, LG Ad Solutions and Raptive were among the launch participants, which PPC Land's report counted as 23 while listing 21 names.

Launched under the MIT licence, the project switched to Apache 2.0 on October 27, 2025, "for additional protections", according to the repository history.

Version 2.5.0 followed on November 22, 2025. Version 3.0.0 arrived on April 22, 2026, with breaking changes: idempotency keys became mandatory on every mutating request, human approval of insertion orders was moved into its own task lifecycle, and schema rules were added to force human review for categories such as housing, lending, employment and pharmaceutical advertising, citing Article 22 of the General Data Protection Regulation (GDPR) and the EU AI Act. The 2.x line was fully deprecated on August 1, 2026, according to the protocol's documentation.

Trusted Match, added to the schemas on March 26, 2026, extended AdCP from planning to serving. It activates pre-negotiated packages at impression time across web, mobile, connected TV (CTV), AI assistants and retail media, with frequency capping that can run outside a DSP. Jochen Schlosser of Adform cited a 50-millisecond service-level target for it. Version 3.1.0 shipped on June 18, 2026 and version 3.2.1 on October 1, 2026.

Even the name has drifted. The original documentation used "Ad Context Protocol"; since the 3.0 release candidate in March 2026 the repository has called it the "Advertising Context Protocol". The acronym is unchanged.

Governance sits with AgenticAdvertising.org, described in the repository as a pending 501(c)(6) nonprofit trade association incorporated in Delaware. Its charter gives four voting classes, brands, agencies, publishers and technology providers, equal board weight regardless of company size.

Why it matters for marketers

The case for AdCP rests on integration cost. Schlosser estimated a custom SSP deals integration at 250 to 350 days of engineering, against about 50 on a shared standard. Across dozens of sellers, that gap decides whether direct deals with smaller publishers are worth pursuing. For publishers, a seller agent is a way to present audience data, packages and price floors to machines without handing them to an intermediary.

The protocol also shifts where buying logic lives. If an agent can discover, book and report through standard calls, some functions that justify DSP fees become portable, an argument Ari Paparo made in July 2025.

Limitations and criticism

Adoption is the central problem. When the protocol launched, Permutive co-founder Joe Root said none of the companies he spoke to was forecasting a single dollar through it, and Anthony Katsur, chief executive of IAB Tech Lab, said "What we don't need is another industry trade group." Fraud researcher Augustine Fou warned that more automation means less transparency. Paparo later questioned whether agents could overcome the problems that stalled programmatic direct: seller resistance to price transparency, scale economics and fragmented data.

Eleven months on, the volumes remain small. Schlosser said in September 2026 that little money flows through either AdCP or its rival, and that the largest platforms had not built AdCP. Google, Amazon, The Trade Desk and Microsoft were absent from the founding roster. Apostra, the company Scope3 became on September 23, 2026, disclosed roughly $1 million in agentic campaign volume, much of it from one advertiser, with about 40 staff after two rounds of layoffs.

Price is another question. DataBeat found agentic buyers clearing at $6.13 CPM against $6.95 for conventional programmatic demand, a 13.4% gap that falls on publisher revenue. Platform access is a fourth limit. Google's developer policies of August 31, 2026 ban "programmatic proxies", naming MCP servers among wrappers that merely re-expose Google Ads capabilities, which constrains any AdCP adapter sitting in front of Google's APIs.

Neutrality is disputed too, since the protocol remains identified with O'Kelley and one company despite membership governance. Schlosser also noted that audience requirements must be fulfilled by each publisher or SSP, adding integration work, and that optimisations such as cost-per-acquisition buying were not yet supported.

What AdCP is not

  • AAMP. IAB Tech Lab's Agentic Advertising Management Protocols, named on February 26, 2026, is the rival framework. It builds on OpenRTB, OpenDirect and the Deals API. Shailley Singh, IAB Tech Lab's chief operating officer, counted 13 functions the two share and rejected the view that they occupy separate layers; Schlosser described AdCP as execution-agnostic and AAMP as an evolution of existing standards.
  • MCP. The Model Context Protocol is a general transport for AI tool calls. AdCP runs on it but adds advertising semantics. As Singh put it, "MCP is a communication substrate; it does not reconcile their advertising semantics."
  • OpenRTB. IAB Tech Lab's real-time bidding protocol handles impression-by-impression auctions in milliseconds. AdCP's media buy tasks operate at planning time, over seconds to days.
  • Contextual advertising. Despite the name, AdCP is not page-content targeting. "Context" is borrowed from MCP and means the information an AI model receives.

Recent developments

Releases are frequent: 3.2.6 and a 3.3.0 beta were both tagged on October 22, 2026. The beta adds an experimental agent identity scheme based on Web Bot Auth and deprecates the earlier RFC 9421 request-signing profiles, which remain functional through 3.x and are due for removal in 4.0. It also stops discovery calls such as get_products from creating accounts or accepting default terms on a buyer's behalf.

Platforms are adding controls around the protocol. PubMatic launched guardrails on August 5, 2026 that validate AdCP traffic and require approvals before agents transact. On the standards side, IAB Tech Lab's AAMP 3.0 draft, with its OpenProposal specification, closed for public comment on October 22, 2026. Schlosser has set out three endings for the contest: one protocol dies, the two merge, or both persist with a translation layer between them. As of October 2026, none has happened.

Timeline

  • July 19, 2025: Brian O'Kelley makes the first commit to the AdCP repository.
  • October 15, 2025: AdCP 2.0.0 launches with six founding members.
  • October 27, 2025: Licence changes from MIT to Apache 2.0; IPR policy added.
  • November 22, 2025: Version 2.5.0 released.
  • December 2025: Scope3 and Magnite run the first agent-to-agent test.
  • January 29, 2026: Prebid.org takes stewardship of the reference sales agent.
  • February 26, 2026: IAB Tech Lab names its rival framework AAMP.
  • March 2026: Repository renames the standard "Advertising Context Protocol" in 3.0 release candidate documentation.
  • March 26, 2026: Trusted Match Protocol added to the schemas.
  • April 22, 2026: AdCP 3.0.0 released.
  • June 18, 2026: Version 3.1.0 released.
  • August 1, 2026: The 2.x line is fully deprecated.
  • August 5, 2026: PubMatic launches guardrails validating AdCP traffic.
  • August 20, 2026: IAB Tech Lab's Shailley Singh counts 13 functions shared by AAMP and AdCP.
  • September 23, 2026: Scope3 renames itself Apostra.
  • October 1, 2026: Version 3.2.1 released.
  • October 22, 2026: Version 3.2.6 and 3.3.0-beta.0 tagged; AAMP 3.0 comment period closes.

Summary

  • Who: Developed by Brian O'Kelley and Scope3 (now Apostra) with founding members Yahoo, PubMatic, Swivel, Triton Digital and Optable; governed by the nonprofit AgenticAdvertising.org, with the reference sales agent maintained by Prebid.org. IAB Tech Lab runs the competing AAMP framework.
  • What: An open, Apache 2.0-licensed protocol of standard tasks and JSON schemas through which AI buyer and seller agents discover inventory, book media, manage creatives, activate audiences and report results, plus a serve-time layer called Trusted Match.
  • When: First committed July 19, 2025; launched October 15, 2025 as version 2.0.0; version 3.0.0 on April 22, 2026; 3.3.0 beta on October 22, 2026.
  • Where: Over MCP and A2A transports between agents run by publishers, SSPs, agencies and DSPs, mapped onto ad servers such as Google Ad Manager, Kevel and Triton.
  • Why: To replace hundreds of custom integrations with one shared interface so AI agents can trade media across platforms, though as of October 2026 little spend runs through it and the largest platforms have not adopted it.