The API era made software talk to software. MCP makes software operable by rented intelligence. Here's what a personalization platform looks like when an agent can run the whole thing.
Every era of software has one sorting question. In 2008 it was "is it in the cloud?" In 2015, "does it have an API?" The 2026 question is "can my agent drive it?" — and most of the martech stack currently answers no.
We think that question is the whole game, so we built AXO the other way around. Every operation in the platform — not a demo subset — is exposed as a tool an agent can call: 130+ of them, over MCP, the open protocol that standardized how software presents its controls to a model. The same surface a human clicks through in the dashboard, an agent drives through tool calls.
Concretely, here is one real session, start to finish. Point an agent at a domain. It crawls the site and maps it — pages, taxonomy, templates. It proposes personalization zones and stages them as drafts. It drafts copy variants for two behavioral segments, attaches imagery, and wires the placements — staged behind review, nothing live yet. You describe how the business sells in one sentence; it proposes a probability-weighted pipeline. A human reads the staged work and approves it, and the tag starts deciding. A week later the agent reads the holdout scorecard, reports which variant is beating control, and suggests what to kill.
Every step of that is a tool call. No step required our UI. And the interesting part isn't any single tool — it's that the tools compose into workflows we never designed. That's the difference between agent tooling and an AI feature: a feature does what the vendor imagined. A tool surface does what your agent can figure out.
"An agent can operate everything" sounds like a liability until you look at where the writes land. Customer-visible changes stage as drafts a human promotes. Pipeline stage changes arrive as append-only suggestions with reasoning attached, not silent edits. Tokens are scoped, actions are logged. And the live per-visitor decision stays deterministic — the agent configures the machine; it doesn't improvise inside it. The design question for agentic software isn't "what can the agent do?" It's "what happens when nobody's watching?" — and the answer needs to be boring.
The MCP server is public — hosted, no local install, authenticated with scoped tokens issued from the dashboard, with the same operations available over REST. We built it because we think the agent-operable version of every software category wins, and we'd rather be that than compete with it.
If you have Claude — or any MCP-capable agent — you can run a personalization program the way you'd run a codebase: describe intent, review the staged diff, ship. The vendors still routing you through their chat box are answering the 2015 question. Ask them the 2026 one.
The full personalization loop: crawl and map a site, propose and create zones, build segments, draft and stage content variants, attach generated imagery, configure triggers and lead forms, manage a sales pipeline, export audiences, and read analytics, holdout lift, and revenue reports. AXO exposes 130+ MCP tools — the same operations available in the dashboard, callable by any MCP-capable agent.
It is safe when the write path is designed for it. In AXO, customer-visible changes stage as drafts a human promotes, pipeline changes arrive as append-only suggestions rather than silent edits, agent tokens are scoped, and actions are logged. The live per-visitor decision is deterministic and untouched by the agent at runtime — the agent configures the system, it does not improvise in production.
An embedded AI assistant is the vendor's agent operating the vendor's product from inside its own UI — it does what the vendor imagined. An MCP tool surface lets your agent operate the product from outside and compose it with the rest of your stack, building workflows the vendor never designed. The test: can an agent you control do real work end to end through the product's public tools?