AI News AI资讯 10d ago Updated 10d ago 更新于 10天前 47

Okta targets AI agent token costs with MCP scoping Okta 通过 MCP 范围限定降低 AI 智能体 Token 成本

Okta proposes identity-scoped Model Context Protocol (MCP) tool lists to reduce AI agent token costs caused by "tool tax" — the overhead of including all tool schemas in every model prompt Filtering tools before prompt construction based on user/agent permissions can reduce visible tools by over 90% in some scenarios, proportionally cutting schema token costs The approach applies least-privilege access at the tool level rather than the MCP server level, with scope checks at both prompt assembly Okta提出身份范围限定MCP工具列表方案,通过权限过滤减少AI agent每次调用的"工具税"token开销 内部建模显示在特定权限场景下可减少90%以上的可见工具数量,工具schema成本近似线性下降 方案在提示构建阶段和运行时执行阶段进行双重作用域检查,实现工具级最小权限访问 与网关层支出控制不同,身份权限控制从源头减少暴露工具,而非事后计量限制

65
Hot 热度
70
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Okta proposes identity-scoped Model Context Protocol (MCP) tool lists to reduce AI agent token costs caused by "tool tax" — the overhead of including all tool schemas in every model prompt
  • Filtering tools before prompt construction based on user/agent permissions can reduce visible tools by over 90% in some scenarios, proportionally cutting schema token costs
  • The approach applies least-privilege access at the tool level rather than the MCP server level, with scope checks at both prompt assembly and runtime execution
  • Identity-based scoping complements but differs from gateway controls, which meter and cap spending after tokens are already consumed
  • Reducing tool visibility also shrinks the attack surface for compromised identities by limiting available actions

Why It Matters

As AI agents increasingly rely on MCP servers to connect to tools and data sources, the cumulative token cost of transmitting full tool catalogs with every model call becomes a significant and often overlooked expense at scale. This approach reframes cost control as an identity governance problem rather than a post-hoc metering challenge, offering a dual benefit of reduced spend and improved security posture for enterprise AI deployments.

Technical Details

  • Tool Tax Problem: Every MCP model call includes schemas, names, descriptions, and parameters for all tools exposed by a server, consuming tokens even for tools the agent will never use — and this cost is incurred before any authorization check occurs
  • Identity-Scoped Filtering: Okta maps MCP server tools to OAuth scopes, then filters the tool catalog per identity using assigned permissions, returning only the scoped tool set to the agent rather than the full server catalogue
  • Two-Layer Scope Check: The first filter occurs during tool list assembly for the agent prompt; a second check runs at runtime before tool execution, enforcing least-privilege at both stages
  • Internal Modeling Methodology: Okta modeled a single MCP client against enterprise tool catalogs, mapping tools to OAuth scopes and defining representative user segments (helpdesk read-only, operators, app admins, super admins) weighted by assumed monthly traffic share
  • Cost Reduction Claims: Some permission scenarios reduced visible tools by over 90%, with tool-schema token costs falling proportionally due to the near-linear relationship between tool count and schema token consumption per prompt

Industry Insight

  • Enterprises should treat MCP tool visibility as both a cost and security concern — unscoped tool catalogs create unnecessary token spend and expand the blast radius of compromised agent identities, making identity governance a critical layer in agentic architecture
  • Gateway-based spending controls alone are insufficient for cost optimization; combining them with identity-level tool scoping addresses the root cause (excessive tool exposure) rather than merely capping downstream consumption
  • As MCP adoption grows, tool-level granularity in access control will likely become a standard expectation, pushing vendors to expose per-tool permission metadata and prompting the emergence of identity-aware MCP middleware

TL;DR

  • Okta提出身份范围限定MCP工具列表方案,通过权限过滤减少AI agent每次调用的"工具税"token开销
  • 内部建模显示在特定权限场景下可减少90%以上的可见工具数量,工具schema成本近似线性下降
  • 方案在提示构建阶段和运行时执行阶段进行双重作用域检查,实现工具级最小权限访问
  • 与网关层支出控制不同,身份权限控制从源头减少暴露工具,而非事后计量限制

为什么值得看

这篇文章揭示了AI agent规模化部署中的关键成本瓶颈——"工具税"问题,为从业者提供了可落地的优化思路。Okta的方案将身份治理与成本控制在同一层实现,兼顾安全与效率,对构建企业级agent系统具有直接参考价值。

技术解析

  • 工具税(Tool Tax)机制:每次模型调用都会将MCP服务器暴露的所有工具的schema、名称、描述和参数包含在prompt中,即使agent永远不会调用某些工具,这些token开销也会累积。成本在模型决策前就已产生,运行时拒绝无法回收已消耗的token。
  • 身份过滤架构:Okta将连接权限从整个MCP服务器粒度细化到单个工具粒度,管理员在Okta仪表板配置特定身份可使用的工具,系统返回作用域限定后的工具集而非完整目录。模型每次接收的prompt仅包含该身份授权的工具列表。
  • 双重作用域检查:第一层在构建agent prompt时过滤工具列表,第二层在agent尝试执行工具调用时再次验证作用域。这种设计确保agent只能感知和使用被明确授权的资源、数据库和工具。
  • 建模方法论:Okta将MCP Server工具映射到OAuth scopes,定义代表用户角色(如helpdesk只读用户、helpdesk操作员、应用管理员、品牌/邮件管理员、超级管理员),按假设的月流量权重计算工具数量减少比例,得出某些场景可减少90%以上可见工具的结论。

行业启示

  • 成本优化应从源头设计:企业构建agent系统时应将身份权限控制前置到工具暴露层,而非依赖网关层的事后计量和限制,这能显著降低大规模部署的token成本。
  • 安全与成本控制的统一架构:身份治理工具可同时解决安全访问控制和成本优化两个问题,减少攻击面的同时也减少了不必要的token消耗,实现安全与效率的双赢。
  • MCP生态需要标准化权限模型:随着MCP成为agent连接工具和数据的标准协议,行业需要建立更细粒度的工具级权限标准和映射机制,以支持企业级安全合规和成本管控需求。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent Security 安全 LLM 大模型