The tools aren't a demo subset. They span the operations the product actually runs on, granular enough to chain: read a segment, draft a variant, stage a placement, check the lift.
The agent answers from live data, drafts against your real zones and segments, and stages anything destructive for review. The human stays in the loop, not out of it.
Point any MCP-capable client at AXO's hosted MCP endpoint and authenticate with an MCP token generated from Settings → MCP Tokens in the dashboard. There is nothing to install; a local server package is also available if you prefer to run it yourself.
The tool surface spans the full personalization loop: reading analytics and holdout lift, querying segments, drafting variants from campaign briefs, modifying placements, firing triggers, and managing profiles. The same operations a person performs in the AXO app, an agent performs as tool calls — 130+ of them.
Scoping and review gates. MCP tokens are issued per-site with role-gated permissions, sensitive write paths require elevated roles, and drafted experiences can be staged for human review before anything is published. You decide where the agent acts autonomously and where it asks first — and the live decision path stays deterministic regardless of what the agent does upstream.
Yes. Every capability exposed over MCP is also reachable over REST with API keys, so scripted integrations and agents share one surface. MCP is the natural fit for conversational agents like Claude; REST fits pipelines and custom code.