Okta is going after the token cost of AI agents with a scoping layer for the Model Context Protocol. Every model call can carry schemas, names and parameters for every tool an MCP server exposes, a bill Okta calls the tool tax, paid even for tools the agent never invokes.
Its fix filters that list through identity permissions before the prompt is built. Internal modeling showed some permission scenarios hiding more than 90% of visible tools, with tool-schema costs falling by roughly the same share, though Okta did not publish absolute token or dollar figures.
The company argues the design tightens security as well. An agent that never sees tools outside its authorization scope cannot attempt to use them, and Okta re-checks scope at runtime before a call executes. It frames the capability inside a blueprint for the secure agentic enterprise.
The overhead compounds at scale, Okta says, because prompt cost repeats on every turn for every active user of a widely used MCP server. Filtering before the prompt is built attacks both the tool-count problem and the user-count problem at once.
