Every vendor now claims AI. The checkable claim is an MCP server — and the delta between "has one" and "an agent can run the product" is where the 2026 martech evaluation actually happens.
Two years after Anthropic introduced the Model Context Protocol, the martech version of the question has stopped being "what is MCP?" and become "which of my tools can my agent actually drive?" It's the right question, and the answers are uneven in ways vendor marketing works hard to blur.
"We have an MCP server" describes three very different products. At the bottom: read-only reporting servers — a handful of tools that fetch metrics, so your agent can summarize a dashboard it cannot act on. In the middle: partial surfaces — real read tools plus a curated sliver of writes, enough for a demo, not enough to run a workflow without a human finishing it in the UI. At the top: operational surfaces, where the product's actual capabilities are tools, writes included, with scoped auth and reads to verify every action. Only the top tier changes how your team works; the other two are API docs wearing a costume.
The test is a workflow, not a tool count alone. Pick something real — "find the underperforming segment, draft two variants, stage them for review, report holdout lift in a week" — and ask whether the agent completes it end to end through the tool surface. Coverage, composability, governance, verifiability: if any of the four is missing, the workflow dead-ends at a human clicking through a UI, which is the thing the agent was supposed to absorb.
As of mid-2026 the experimentation and personalization category is mostly unagentic. Adobe Target publishes no public MCP server — your agents cannot drive it. The legacy testing vendors are in similar shape. The notable exceptions run in opposite directions: Mutiny's relaunched GTM-asset product is MCP-native for generating collateral, though it no longer touches your website; and AXO exposes 130+ MCP tools spanning the whole personalization loop — segments, variants, placements, triggers, lead capture, pipeline, holdout measurement — the same surface the app itself uses.
That asymmetry is temporary; every vendor will ship something labeled MCP within a couple of quarters. Which is why the evaluation habit matters more than any current ranking: ask for the tool list, run one real workflow, and check where the writes are. The vendors confident in their surface will hand you a token. The ones who send a slide instead have answered the question too.
Ask for the complete tool list, then attempt one real workflow end to end with an agent — not a demo the vendor drives. Check four properties: coverage (are the product's real operations exposed, including writes), composability (can tools chain into workflows), governance (scoped tokens and review gates), and verifiability (can the agent read back what its writes did). A server that fails any of the four will strand workflows at a human.
A read-only server exposes tools that fetch data — reports, metrics, lists — but none that act. It lets an agent summarize the product, not operate it, so every workflow still ends with a person clicking in the vendor's UI. Many martech MCP claims in 2026 describe read-only servers; the distinction is the first thing to check, because the marketing rarely volunteers it.