Kong

Any REST API can become an MCP server through the AI MCP Server entity in Kong AI Gateway's MCP Traffic Gateway, which also aggregates several MCP servers behind one endpoint for clients such as Claude Desktop and Cursor. AI consumers authenticate through an AI Auth Strategy (key auth or OpenID Connect, with OAuth2 metadata from the MCP spec), and AI Policies handle access control and observability. Agents take authenticated action on the customer's systems through the gateway.

Sourcedeveloper.konghq.com/ai-gateway/mcpread 2026-09-22

Agent to agent requests pass through Kong's A2A Traffic Gateway, which routes them via AI Gateway, extracts task metadata and rewrites agent cards so clients route through the gateway; the agents and their steps run on the customer's own servers. No workflow definition or agent execution of Kong's own is documented, nothing that sequences, branches, retries or routes work across deterministic workflow nodes and autonomous agent steps.

Sourcedeveloper.konghq.com/ai-gateway/a2aread 2026-09-22

RAG runs at the gateway through the AI RAG Injector plugin: content is sent to a configured vector database through the Admin API or decK, and for each request the gateway embeds the prompt, queries the vector database for the most similar content and injects it before the model call, with vector database access limited to the gateway. New customer knowledge enters a persistent index without retraining, and that index scales past the context window and stays queryable. Some of this sits on higher tiers.

Sourcedeveloper.konghq.com/plugins/ai-rag-injectorread 2026-09-22

Safety and data loss policies apply to model traffic automatically: prompt guards, semantic prompt and response guards, PII sanitization, and third party guardrails from AWS, Azure, Google Model Armor and Lakera, plus token based rate limits and per consumer access control on models, MCP servers and agents. These are constraints the customer controls on what an agent may do. No step where a person reviews, approves or must validate before an agent action commits is documented.

Sourcedeveloper.konghq.com/ai-gateway/ai-data-govread 2026-09-22

Access controls are deep and documented: Konnect SSO through OpenID Connect or SAML with IdP attribute mapping, custom teams and roles, Konnect and Dev Portal audit logs, and on the AI Gateway itself AI Auth Strategies (key auth or OpenID Connect) authenticating each AI consumer for models, MCP servers and agents, with FIPS 140-3 support in Kong Gateway. No security attestation is confirmed on Kong's own pages.

Sourcedeveloper.konghq.com/konnect-platform/ssoread 2026-09-22

AI traffic is recorded for the customer's review: logs and LLM metrics viewable in Konnect or exported to any OpenTelemetry tool, a standardized AI audit log format for AI policies (Data Governance page), OpenTelemetry metrics for generative AI, MCP and A2A traffic, and protocol aware metrics and tracing for agent to agent calls. That gives step by step inspection of what agents sent and received, an audit record separate from runtime metrics, and export to the customer's own tools.

Sourcedeveloper.konghq.com/ai-gateway/ai-data-govread 2026-09-22

Kong AI Gateway governs traffic in flight: its documented AI plugins and entities cover proxying, guardrails, rate limiting, caching, RAG injection, MCP and A2A routing and observability, and none persists an agent's state or a user's context across sessions, so there is no memory layer with a stated scope and lifetime.

Sourcekonghq.com/pricingread 2026-09-22

Konnect lets an organization host its control plane in a region it chooses (Australia, EU, Middle East, US, India or Singapore, picked in a region picker), with gateway services, routes, consumers and other objects kept in that geo. Dedicated Cloud Gateways run in named AWS regions, and data planes can be self hosted in hybrid, DB-less or traditional on premises topologies.

Sourcedeveloper.konghq.com/konnect-platform/geosread 2026-09-22

Prebuilt assets for AI traffic come as plugins (guardrails, prompt templates and decorators, rate limiting, caching, RAG injection) that configure or constrain the customer's agents rather than do work themselves, and the MCP Registry catalogs the customer's own MCP servers. No ready made workflows, templates or role specific agents for a buyer to adopt are published.

Sourcekonghq.com/pricingread 2026-09-22

Every model call, MCP tool call and A2A call that passes through Kong AI Gateway is started by the caller and then routed, secured and observed in flight. No schedule, event or webhook that starts an agent run is documented; the gateway acts only on requests a client has already sent.

Sourcedeveloper.konghq.com/ai-gateway/a2aread 2026-09-22

The customer chooses: Kong AI Gateway offers a universal LLM API across providers, with AI Proxy Advanced balancing across models the customer configures and LLM access control and authentication per consumer. The pricing page lists the universal LLM API, token based rate limiting and semantic caching on Plus and bills per unique model proxied.

Sourcekonghq.com/pricingread 2026-09-22

Kong publishes OpenAPI specifications for the Gateway Admin API, the Konnect AI Gateway API and the Konnect AI Catalog API, supports declarative configuration through decK, Terraform and the Kong Operator, and ships a Kong Konnect MCP Server so AI clients can manage Konnect through MCP tools.

Sourcedeveloper.konghq.com/llms.txtread 2026-09-22

Responses can be scored by a dedicated judge model through Kong's AI LLM as Judge plugin, which evaluates prompt and response pairs and gives each response a score from 1 to 100, with a configurable evaluator prompt, conversation history depth and sampling rate. Paired with AI Proxy Advanced's accuracy balancer, it compares responses from at least two models on live traffic, so customers get a readable score they can compare on their own traffic.

Sourcedeveloper.konghq.com/plugins/ai-llm-as-judgeread 2026-09-22

Browser and computer control sit outside what Kong AI Gateway does: it routes and governs model, MCP and agent traffic, and no browser, desktop or remote computer session it runs for an agent is documented. No headless fetch or third party engine is wired in either.

Sourcekonghq.com/products/kong-ai-gatewayread 2026-09-22