Okta Targets AI Agent Token Costs with MCP Tool Scoping
Okta introduces identity-based scoping for Model Context Protocol (MCP) tools to cut prompt overhead and enhance agent security.
Stock photo for illustration only, not from the actual event
- Okta highlights token waste from exposed AI agent tools.
- Proposes identity-scoped filtering for MCP tool lists before model calls.
- Internal models show over 90% reduction in visible tools in certain scenarios.
- Combines cost reduction with least-privilege security controls.
Okta has outlined a new approach to cutting token expenses for AI agents by scoping Model Context Protocol (MCP) tool lists according to user and agent identities, aiming to eliminate overhead costs incurred before a model even processes a prompt.
In standard AI agent operations, every model call includes schemas, names, descriptions, and parameters for all connected tools exposed by an MCP server. Okta refers to this prompt overhead as the “tool tax,” representing tokens consumed as models evaluate tools they may never actually invoke.
This financial burden compounds with every active user and expanding tool catalogues. Furthermore, there is an access-control risk: if an agent can view tools outside its authorized scope, it may attempt to use them. While runtime rejections can block execution, the model has already ingested the tool definition and consumed tokens to process it.
Okta’s proposed solution filters the tool list within the Okta dashboard using assigned identity permissions before it reaches the model. The agent receives a streamlined list, and a secondary check runs at execution time to enforce strict least-privilege access at the tool level.
Stock photo for illustration only, not from the actual event
Scoping Model Context Protocol (MCP) tools at the identity layer bridges a critical gap between enterprise cybersecurity and modern AI orchestration. Historically, developers exposed entire tool servers to agents without granular user-level partitioning. Filtering unneeded tools at the ingestion stage not only slashes API token overhead but fundamentally shrinks the potential blast radius if an agent is compromised.
Based on internal modeling using Okta product data and vendor documentation—incorporating user segments such as read-only support, app administrators, and super administrators weighted by traffic share—Okta found that certain permission setups removed more than 90% of visible tools. Because each tool contributes its parameter schema to every prompt, token costs tracked tool counts nearly linearly.
"Cost control for agents is best provided using identity governance tools that offer more granular control and precision without disrupting business processes."
Paul Webber, Principal Cybersecurity Industry Analyst at Software Analyst Cyber Research
Okta notes that actual results will depend on specific tool catalogues, permission layouts, chosen models, and pricing structures. This identity-layer filtration complements traditional API gateways, which typically manage rate-limiting and spending caps after token expenditure has already accrued.
Source: AI News
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment