Okta proposes identity-based filtering to cut MCP tool costs
Okta says filtering Model Context Protocol (MCP) tools according to an agent’s identity could reduce the token cost of AI model calls. MCP servers provide tools through schemas containing names, descriptions and parameters, and that information may be added to every prompt, including tools the agent will not use. Okta refers to this recurring prompt overhead as the “tool tax”.
The proposed approach uses permissions assigned to an agent and its associated user to create a narrower tool list before the prompt is built. Okta says internal modelling found that some permission scenarios reduced the number of visible tools by more than 90%, with schema-related token costs falling by roughly the same proportion. The company did not provide absolute token or dollar savings.
Administrators would configure tool permissions in the Okta dashboard, while a second check would take place before execution. The company says this applies least-privilege access at the individual-tool level and limits the actions available to a compromised identity. Its modelling used OAuth scopes, representative user groups and assumed traffic shares, rather than customer data. Results can vary with the tool catalogue, permissions, model, schema size, request volume and pricing. Okta distinguishes this filtering from gateways, which can meter tokens, spending and traffic after requests enter the system.
Written by Sakhanda Wire AI from reporting by AI News. Facts, quotes and full context are in the original article below.