Adform's Solutions Engineering team this week released 29 agentic skills that let AI systems query the Adform FLOW demand-side platform in plain language, according to a GitHub repository the company published under the name agentic-skills. The skills run through the Adform GraphQL Model Context Protocol server and cover reporting, forecasting, audience discovery, and campaign and creative audits, and every one of them is read-only.

Your Feed Has a Viewability Problem

If nobody's reading it, did it ever really load? Add PPC Land as your preferred source in Google Search and actually see the stories that matter.

Fix My Sources

What the skills actually do

Each skill in the release is a single markdown file. According to the repository's README, that file "defines the skill's purpose, illustrative GraphQL queries, usage constraints, and presentation guidance." There is no executable code bundled with most of them, no separate binary, and no write access to Adform FLOW. An AI agent connected to the Adform GraphQL MCP server reads the skill file, learns which GraphQL queries answer a given question, and returns an answer or a report. It cannot push a change back into the account.

That constraint is stated explicitly and repeatedly across the individual skill files. The taxonomy governance skill, one of the 29, describes itself in its own frontmatter as validating "entity names against a convention pattern" and flagging "missing or wrong label taxonomy across campaigns." Directly underneath that description sits a one-line qualifier: "Read-only - it reports; the trafficker fixes." The same pattern repeats through the collection. Skills forecast, they audit, they summarize, and they flag. None of them execute a change inside the platform.

The taxonomy governance skill offers a useful illustration of how granular the read-only design gets. It pulls a list of campaigns for a given advertiser, then applies a naming convention as a regular expression on the client side. The example pattern given is ^[A-Za-z]+_\d{4}_Q[1-4]_[A-Za-z]+$, matching a structure such as Brand_Year_Quarter_Objective. Any campaign name that fails to match gets flagged, as does any case-insensitive duplicate found within the same advertiser's scope. A second half of the same skill checks label taxonomy: it retrieves the label groups defined for an advertiser, such as Market, Product, or Funnel Stage, then compares those against what has actually been applied to each campaign, flagging any campaign missing a required group. The output format is prescribed too. According to the skill file, the presentation should consist of two compliance tables, one for naming and one for labels, and it should "lead with the non-compliant count and the items that need the most immediate attention." Every flagged row, the file states, "should be a clear action item for the trafficking or operations owner."

The connection tooling section of that same file gives a small technical detail worth noting for anyone building on top of this: calls to the GraphQL endpoint should be spaced roughly one to two seconds apart. That is a rate-limiting consideration built directly into the skill's operating instructions rather than left for a developer to discover through trial and error.

Scope of the 29 skills

The GitHub repository's skills directory lists folders covering a wide functional range. Based on the repository structure, these include audience discovery, bid landscape analysis, budget risk monitoring, campaign management, campaign performance, campaign structure, channel conflict auditing, creative discrepancy auditing, deal health checks, delivery health monitoring, entity browsing, frequency cap auditing, geo reference lookups, inventory availability forecasting, inventory deals, line items, media plan review, media plans, pacing checks, past traffic analysis, reach forecasting, reporting, segment governance, stale entity auditing, stats performance, targeting segments, taxonomy governance, tracking tags, and viewability auditing.

Suren Silva, VP of Global Solutions Engineering at Adform, announced the release on LinkedIn, describing it as covering "reporting, forecasting, audience discovery and campaign/creative audits (read-only) on top of the Adform #MCP." Silva credited three named colleagues by name for the work: Marcel Ehrlitzer, Hans Jirschik, and Maja Sokołowska, along with what he described as the Product and Engineering teams responsible for building and enhancing what he called the MCP Gateway. His post ended with a note that more skills are planned: "p.s. A LOT more skills to come."

The repository's own commit history bears that description out. The project has accumulated 26 commits under a public MIT license, with three listed contributors on GitHub: Suren Silva, Hans-Ferdinand Jirschik, and Marcel Ehrlitzer. The most recent commit, a pull request merge titled "fix: sync audience-discovery skill across all dist locations and rebuild," touched the audience discovery skill specifically, suggesting the release is still being actively refined rather than treated as a finished, static drop.

How the technical distribution works

The repository organizes its publishable output under two parallel folders, according to the README. One, dist/claude/, contains what the documentation calls "packaged skill plugin for Claude (Anthropic)." The other, dist/generic/, holds what the same documentation describes as "platform-agnostic skills following the agentskills.io standard." That split matters for anyone trying to use these skills outside Anthropic's ecosystem specifically. A trafficker or analyst running Claude Code, Anthropic's command-line coding tool, gets a packaged plugin they can install directly. Anyone using a different agentic AI system is pointed toward the generic, platform-agnostic version instead.

The recommended installation path for Claude Code involves two commands run from the command line: claude plugin marketplace add adform/agentic-skills, followed by /plugin install adform-agentic-skills@adform-agentic-skills. The README also lists a fallback path, uploading a zip file named adform-agentic-skills.zip through a plugin interface, for situations where the marketplace command does not work as expected.

Before any of this functions, the underlying agentic AI system has to be connected to the Adform GraphQL MCP server itself. The README points to a separate page, hosted at solutions.adform.com, with setup instructions for that connection. The skills sit on top of that server; they do not replace the need to configure it.

Context: MCP adoption across the industry

The technical foundation underneath Adform's release, the Model Context Protocol, did not originate inside advertising technology. Anthropic developed and released the open standard, and later donated it to the Linux Foundation, as PPC Land has documented in earlier coverage of MCP adoption across the sector. Over roughly the past eighteen months, the protocol has moved from a niche developer tool into what amounts to connective tissue across programmatic infrastructure, with implementations spanning AdRoll, PubMatic, FreeWheel, Yahoo DSP, and DoubleVerify.

Adform's own MCP work did not begin with this release. Google explored an MCP server for its Ads API in July 2025before releasing an open-source version that October, and a wave of platforms followed a similar path through late 2025 and into 2026. Microsoft's Clarity MCP server launched in June 2025 for analytics queries. Yahoo DSP rolled out agentic capabilities including a Model Context Protocol-based campaign activation agent on January 6, 2026, letting external agents connect to the platform under what Yahoo called its "Yours, Mine, and Ours" framework.

What distinguishes Adform's 29-skill release from many of the platform announcements PPC Land has covered in this space is the explicit, repeated emphasis on read-only access. Where Meta opened its ad system to Claude and ChatGPT in April 2026 with write access from day one, allowing AI agents to create and edit campaigns directly, Adform's skills stop at the reporting and flagging stage. Google's Ads API MCP server, released as open source in October 2025, launched read-only as well, and Amazon Ads followed the same conservative posture when its own MCP server entered open beta. Microsoft's expanded Advertising MCP server, announced at Cannes Lions in June 2026, likewise opened in pilot with read-only access, with Rukmini Iyer, Corporate Vice President of Microsoft AI, framing the phased approach as a deliberate choice rather than a technical limitation. Adform's release sits inside that same conservative bracket: an agent can tell a trafficker their naming conventions are broken, but it cannot rename the campaign itself.

The scale of Adform's release also invites direct comparison with a similar effort elsewhere in the industry. Kochava's StationOne platform, which opened to public beta on March 25, 2026, shipped with an IAB Tech Lab Agentic Advertising Management Protocols workspace containing 19 specialized agentic skills organized across 8 functional areas, all running through IAB Tech Lab's reference implementation MCP server. Adform's 29 skills, built for a single company's own DSP rather than for an industry-wide standard, represent a larger individual set, though the two releases are not built on the same underlying protocol layer and are not directly interchangeable.

Why this matters to marketers

For agencies and in-house teams that already run campaigns through Adform FLOW, the practical change is narrower than the words "AI agent" might suggest. Nothing about campaign delivery, bidding, or budget pacing changes as a direct result of this release. What changes is how quickly certain diagnostic and compliance questions can be answered. A media buyer who previously had to open FLOW, navigate to a specific report, filter it, and cross-reference a separate label configuration screen can instead ask a connected AI agent the same question in natural language and receive a structured answer.

The taxonomy governance skill is a concrete example of the kind of operational friction this could reduce. Naming convention drift and inconsistent label application are common, unglamorous problems inside large advertiser accounts running many campaigns across many managers. Catching them manually requires someone to pull an entity list, apply a naming pattern by eye or by spreadsheet formula, and cross-check label groups campaign by campaign. According to the skill file itself, that entire process can now be delegated to an agent that returns two compliance tables, with the non-compliant items surfaced first.

The read-only boundary carries a governance implication too, one that PPC Land's coverage of agentic advertising has returned to repeatedly. Ari Paparo, an ad tech veteran, published an analysis in November 2025 raising concerns about media buying applications of agentic protocols more broadly, while separately praising the more structured creative specification work happening in the space. Anthony Katsur, CEO of IAB Tech Lab, has been publicly skeptical of protocol proliferation in the sector, though he has also acknowledged that agentic capability represents something more substantive than a passing trend. A release built entirely around reporting and flagging, with no execution capability at all, sidesteps a large portion of that governance debate. There is no question of an AI agent overspending a budget or mistargeting a campaign here, because the skills are architecturally incapable of taking either action.

That does not mean the release is without judgment calls left to the human operator. The taxonomy skill's own documentation states plainly that it will "suggest a corrected name where the intent is clear from the existing name," which means the agent is making an inference about what a human trafficker originally meant. Whether that inference is reliable across edge cases, unusual campaign names, or advertiser-specific naming quirks is not something the skill file itself addresses, and it is the kind of question that typically only becomes visible once a tool like this is running against real account data at scale.

PPC Land has tracked the Model Context Protocol's expansion across advertising infrastructure closely over the past year. Six companies, including PubMatic, Scope3, and Yahoo, launched the Ad Context Protocol as an advertising-specific extension of MCP in October 2025, and that launch drew immediate industry debate over whether the sector needed another standard. Displayce introduced three MCP-connected agents for digital out-of-home advertising at Cannes Lions in June 2026, describing the category as Agentic DOOH. DoubleVerify built Model Context Protocol support into a product it calls the Neura Insight Agent, accessible through Anthropic's Claude for natural language queries against media quality data.

Whether Adform's 29 skills see meaningful adoption will depend on factors the repository itself cannot answer: how many agencies actually connect their AI tooling to the Adform GraphQL MCP server, how reliable the read-only outputs prove to be against messy real-world account data, and whether Adform's promise of "a lot more skills to come" translates into a steady release cadence or tapers off. The commit history shows active maintenance as of this writing, with the audience discovery skill having been synchronized and rebuilt across distribution locations in the most recent update.

Timeline

  • Adform's Solutions Engineering team builds and publishes the adform-agentic-skills repository on GitHub, structured with separate distribution paths for Claude and platform-agnostic AI systems
  • The repository accumulates 26 commits under an MIT license, with three named contributors
  • Suren Silva, VP of Global Solutions Engineering at Adform, publishes a LinkedIn post announcing the shipment of 29 agentic skills covering reporting, forecasting, audience discovery, and campaign and creative audits
  • The most recent commit to the repository, a pull request merge, synchronizes the audience discovery skill across all distribution locations and triggers a rebuild

Summary

Who: Adform's Solutions Engineering team, credited by name to Suren Silva, Marcel Ehrlitzer, Hans Jirschik, and Maja Sokołowska, published the release. Adform operates the FLOW demand-side platform from its headquarters in Copenhagen, Denmark.

What: A public GitHub repository containing 29 read-only agentic skills, each a markdown file defining a specific query or audit function, that let AI systems connected through the Adform GraphQL Model Context Protocol server answer natural language questions about reporting, forecasting, audience discovery, and campaign or creative compliance inside Adform FLOW. None of the skills can modify a live campaign.

When: Suren Silva announced the release on LinkedIn, describing the skills as having just been shipped. The GitHub repository shows 26 commits accumulated over time, with the most recent commit made the day before the announcement.

Where: The skills are distributed through a public GitHub repository named adform-agentic-skills, with installation paths for Claude Code specifically and for platform-agnostic AI systems following the agentskills.io standard.

Why: The release extends a broader pattern PPC Land has tracked across advertising infrastructure, in which demand-side and supply-side platforms build Model Context Protocol connections to let AI agents query campaign data conversationally. Adform's read-only design places it in the more conservative bracket of that trend, alongside similar launch postures from Google, Amazon, and Microsoft, and distinct from write-access implementations such as Meta's Ads AI Connectors.