Okta slashes AI agent costs with fine-grained MCP tool access
AI agents that rely on Model Context Protocol (MCP) servers can now avoid a hidden tax on every prompt. Okta says its identity-scoped MCP tool lists can cut the tokens consumed by “tool tax”—the overhead of schemas, names and descriptions sent to the model on every turn. By filtering tools before they reach the prompt, the company claims some permission scenarios reduce visible tools by more than 90%, trimming an equivalent share of schema costs. Okta’s evidence comes from internal modelling using OAuth scopes and representative roles rather than live customer data.
A costly side-effect of broad tool exposure
MCP servers act as bridges between AI agents and external tools or data, exposing everything from Google Workspace integrations to internal APIs. Each exposed tool arrives in the model’s prompt as a schema plus metadata, inflating token usage even when the agent never calls it. When many tools are visible to many users, the cumulative cost compounds quickly. Okta frames the issue as both a tool-count and user-count problem—each active user incurs the prompt overhead on every model call.
Identity-first filtering at the edge
Okta’s approach narrows access from whole servers to individual tools, assigning least-privilege permissions at the tool level. Administrators configure which tools an identity may use in the Okta dashboard; the system then returns a scoped list instead of the server’s full catalogue. The model receives only the authorised subset, reducing schema cost before the prompt is built. A runtime check still blocks unauthorised calls, but the tokens for excluded tools are never consumed in the first place.
Why it matters
For enterprises scaling AI agents across large MCP catalogues, token savings translate directly into lower inference bills and faster responses. Okta’s model-driven estimates suggest dramatic reductions in overhead, though absolute figures remain unpublished. More broadly, the feature reinforces the need for fine-grained, identity-centric access control in agentic systems, where every unnecessary schema is both a security risk and a cost centre. Expect identity platforms to keep tightening the screws on MCP economics as adoption grows.
Source: AI News. AI-assisted editorial synthesis — TechnoExpress.