AI 模型和平台
Kong 新增 AI 对 Konnect 的成本、信誉和可观察性控制
在 Google 上的首选来源中添加 Unite.AI
Kong 推出 AI Management in Kong Konnect,2026年11月4日，随着企业将agents移向生产，增加了对AI支出，资信，管理资源发现和可观察性的控制。
成本管理，认证，上下文和基本可观察性组件与Kong AI Gateway一起出现在Konnect上，而不是作为单独的加载，因此已经运行该网关的团队可以不用另外购买即可使用。 更宽的发布范围还适用于模型的发现，模型背景协议(MCP)服务器和agents.
Kong正在解决四个操作漏洞：无法分配到团队或应用程序的支出，嵌入agent代码或提示的证书，API反复包裹为agent工具，以及没有agent活动的综合账户。 当一个任务跨越多个模型、MCP服务器、API、工具和其他agents时，这些空白就变得难以管理。 将控制放入连接路径可以使政策和使用数据遵循这些相互作用，而不是由每个应用团队重建。
成本分配和运行时间优化
AI Cost Management将网关消费数据与请求身份和上下文相结合，将支出归属于业务单位，工程团队，应用程序，agents，个人用户，客户或租户。 其价格目录集中了供应商的代币和支出率，允许超越谈判安排。 与第三方AI工具的集成将归属扩展至编码带，copilots和软件开发工作流程。
各小组可以针对这些归属群体制定预算，并随着消费接近阈值而得到警报。 Konnect概述了新增软硬预算选项，以及一个仪表板，按团队、利用或存放单位比较支出与预算上限。
优化通过智能模型路由，语义缓存，即时缓存，即时压缩和推理控制在基础设施层应用。 Kong 称的潜在好处因工作量而异：语义缓存特别适合重复请求或类似请求，而当昂贵的模型处理低成本替代品能够完成的任务时，路由可能更重要。 机会是寻找更多效率和衡量最优化节余的一个即将到来的特点，它被标记为即将到来，而不是包括在最初的发布中。
Agent 交通记录
Token Vault 发布在Kong AI Gateway 2.3中带有每个用户三脚OAuth，并包含所有Kong Gateway订阅。 一个用户和agent认证到MCP服务器代理，使用公司身份提供者证书。 进入、记录、限制费率和其他网关政策在上载上游证书之前得到评价。
在第一次使用时，网关可以通过MCP URL诱导或合成工具返回短寿命的入学链接。 用户完成上游服务正常的OAuth同意流，之后网关会用客户的密钥加密由此产生的刷新令牌，并将加密的信封发送给Kong IDIS存储。
在后来授权的呼叫中，网关用从客户的秘密管理器读来的私人密钥解密了内存中的证书，将其插入上游请求中，并刷新和重新加密过期的令牌。 Kong 称 agent接收回复时没有看到符号。 Konnect 存储加密的令牌，但缺少打开它们所需的密钥；即使上游证书尚未取消，删除用户的政策访问也立即生效。
管理下的发现和操作可见度
Kong AI Register涵盖商业，自办和定制型号，MCP服务器，以及内部和第三方的agents. 开发者和agents可以发现受制裁的资源，而不仅仅依赖于每个agent的硬编码连接。 仅凭登记不能提供访问：AI Gateway提供认证、授权、政策执行、交通管理和运行时的可观察性，在控制中设有费率限制和安全护栏。
上下文Mesh从现有的API和MCP服务器中生成受管的MCP服务器，将多个源集成复合材料，并可以将由此产生的服务器部署到AI Gateway. 使用代码模式，agent搜索可用的API表面，检索相关的计划并执行操作，而不是加载每个工具定义。 它还可以用环路，条件和错误处理来构成操作。 Kong已经于2026年9月30日将背景管理作为AI Management的一部分普遍提供；而之前提供的是新包装的背景。
AI Observation目前提供庄园概览，预建和自定义的仪表板，用于临时查询和连接日志的探索器。 Dashboards按型号，agent,MCP工具，团队或环境分解流量和支出。 共享观点受角色访问控制，而OpenTeleometter和度量衡输出方则允许遥测为现有可观察性系统提供素材。
2026年11月下旬或12月初，预定在agent步骤、工具呼叫和模式呼叫之间进行高级会话和回合跟踪。 计划的工作流程允许各小组检查每个回合的投入、产出、标志、成本和耐久性，以及产生结果的政策检查和工具呼叫。
发布将网关的作用扩展至传递或屏蔽个人AI请求之外：它成为agent活动附加身份，预算和工具访问政策的共享层。 现有的Konnect网关客户获得该软件包内的核心成本、证书和发现控制，而整个会话的审计线索和优化节约的证据则属于推出的下一阶段。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
Kong Adds AI Cost, Credential and Observability Controls to Konnect
Add Unite.AI to your preferred sources on Google
Kong introduced AI Management in Kong Konnect on November 4, 2026, adding controls for AI spending, credentials, governed resource discovery and observability as enterprises move agents into production.
The cost-management, credential, context and basic observability components come with Kong AI Gateway on Konnect rather than as separate add-ons, so teams already running that gateway can use them without another purchase. The broader release also introduces governed discovery for models, Model Context Protocol (MCP) servers and agents.
Kong is addressing four operational gaps: spending that cannot be assigned to a team or application, credentials embedded in agent code or prompts, APIs repeatedly wrapped as agent tools, and the absence of a consolidated account of an agent’s activity. Those gaps become harder to manage when one task crosses multiple models, MCP servers, APIs, tools and other agents. Putting controls in the connectivity path allows policy and usage data to follow those interactions instead of being reconstructed by each application team.
Cost Attribution and Runtime Optimization
AI Cost Management combines gateway consumption data with request identity and context to attribute spending to business units, engineering teams, applications, agents, individual users, customers or tenants. Its price catalog centralizes provider token and spending rates and permits overrides for negotiated arrangements. Integrations with third-party AI tools extend attribution to coding harnesses, copilots and software-development workflows.
Teams can set budgets against those attribution groups and receive alerts as consumption approaches thresholds. The Konnect overview adds hard and soft budget options and a dashboard for comparing spending with budget ceilings by team, harness or repository.
Optimization is applied at the infrastructure layer through intelligent model routing, semantic caching, prompt caching, prompt compression and reasoning controls. Kong says the potential benefit varies by workload: semantic caching is particularly suited to repetitive or similar requests, while routing can matter more when expensive models handle tasks that lower-cost alternatives could perform. Opportunities, a forthcoming feature for finding additional efficiencies and measuring optimization savings, is marked as coming soon rather than included in the initial release.
Per-User Credentials for Agent Traffic
Token Vault launched with per-user three-legged OAuth in Kong AI Gateway 2.3 and is included with all Kong Gateway subscriptions. A user and agent authenticate to an MCP server proxy using corporate identity-provider credentials. Access, logging, rate limiting and other gateway policies are evaluated before an upstream credential is loaded.
On first use, the gateway can return a short-lived enrollment link through MCP URL elicitation or a synthetic tool. The user completes the upstream service’s normal OAuth consent flow, after which the gateway encrypts the resulting refresh tokens with the customer’s key and sends the encrypted envelope to Kong Identity for storage.
On later authorized calls, the gateway decrypts the credential in memory with a private key read from the customer’s secrets manager, inserts it into the upstream request, and refreshes and re-encrypts expired tokens. Kong says the agent receives the response without seeing the token. Konnect stores the encrypted tokens but lacks the key needed to open them; removing a user’s policy access takes effect immediately even if the upstream credential has not yet been revoked.
Governed Discovery and Operational Visibility
Kong AI Registry covers commercial, self-hosted and custom models, MCP servers, and internal and third-party agents. Developers and agents can discover sanctioned resources instead of depending only on connections hard-coded into each agent. Registration alone does not grant access: AI Gateway supplies authentication, authorization, policy enforcement, traffic management and observability at runtime, with rate limits and security guardrails among the controls.
Context Mesh generates governed MCP servers from existing APIs and MCP servers, combines multiple sources into composites and can deploy the resulting servers to AI Gateway. With Code Mode, an agent searches the available API surface, retrieves the relevant schema and executes the operation rather than loading every tool definition. It can also compose operations with loops, conditionals and error handling. Kong had already made Context Management generally available as part of AI Management on September 30, 2026; that earlier availability is background to the new packaging.
AI Observability currently provides an estate overview, prebuilt and custom dashboards, an Explorer for ad hoc queries and connectivity logs. Dashboards break down traffic and spending by model, agent, MCP tool, team or environment. Shared views are subject to role-based access controls, while OpenTelemetry and metrics exporters allow the telemetry to feed existing observability systems.
Advanced session-and-turn tracing across agent steps, tool calls and model calls is scheduled for late November or early December 2026. The planned workflow lets teams inspect an individual turn’s inputs, outputs, tokens, cost and latency alongside the policy checks and tool calls that produced the result.
The release extends the gateway’s role beyond passing or blocking individual AI requests: it becomes a shared layer for attaching identity, budgets and tool-access policy to agent activity. Existing Konnect gateway customers receive the core cost, credential and discovery controls within that package, while the full session audit trail and evidence of optimization savings belong to the next stage of the rollout.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.