MCP
Page Agent MCP turns browser control into an agent interface
MCP is attractive because it can let external agent clients ask a browser to perform approved work. For Page Agent, the practical question is how to expose that power without losing user consent, auditability, and product context.
For teams considering Page Agent MCP after an in-page JavaScript pilot or extension rollout.
Start with in-page policy
Before exposing browser actions through MCP, define the in-page action policy: which selectors are allowed, which routes are blocked, which actions require confirmation, and which events must be logged.
Once that policy is stable, MCP can extend the same rules to external clients instead of inventing a second trust model.
- Name the clients allowed to request browser actions.
- Require explicit consent for high-impact tasks.
- Record prompts, planned actions, approvals, and results.
- Keep secrets out of prompts, logs, and visible page content.
When MCP belongs in the first purchase
MCP belongs in the first purchase when the buyer already needs external agent clients, cross-tool coordination, or browser control from a larger automation system. Otherwise, launch the in-page copilot first and add MCP after usage data proves the workflow.
The pricing page keeps Pro annual selected by default. Scale is available for teams that need MCP planning, extension rollout, and heavier governance from day one.
What success looks like
A successful MCP rollout does not feel like magic. It feels like a reliable control surface: the user knows what client asked for action, what the browser will do, what happened, and how to stop or reverse a workflow when needed.
Common questions
Does every Page Agent launch need MCP?
No. MCP is useful for external agent clients, but most product teams should prove the in-page workflow first.
Which plan includes MCP planning?
Scale is the right fit when MCP is required immediately. Pro annual is still the default for an owned SaaS in-page rollout.
How should MCP actions be audited?
Log the requesting client, user confirmation, intended action, result, failure state, and any human handoff.