# Kong AI Gateway

**Canonical:** https://apis.io/providers/kong-ai-gateway/  
**Website:** https://konghq.com/products/kong-ai-gateway  
**APIs profiled:** 21

Kong AI Gateway is the AI-native capability layer built on top of Kong Gateway and managed through Kong Konnect. It exposes a normalized, provider-agnostic LLM API across 16+ providers (OpenAI, Anthropic, Azure AI, Amazon Bedrock, Google Gemini, Vertex AI, Cohere, Hugging Face, Llama, Mistral, xAI, DashScope, Cerebras, Ollama, Databricks, DeepSeek, vLLM), and adds prompt firewalls, PII sanitization, semantic caching, automated RAG injection, token-level observability, per-agent cost allocation, MCP traffic governance, and Agent Gateway support for agent-to-agent (A2A) communication. It is profiled here as a standalone product surface; the parent provider profile lives at github.com/api-evangelist/kong.

## Kin Score — 33.5 / 100 (thin)

Scored 2026-08-20 under rubric 0.12.0. Trend: flat (+0.0 from 33.5).

| Facet | Score |
|---|---|
| Discoverability | 72.2 |
| Contract Quality | 0.0 |
| Governance | 0.0 |
| Contract Governance | 0.0 |
| Operational Transparency | 50.0 |
| Developer Ergonomics | 59.5 |
| Commercial Clarity | 39.5 |
| Access Clarity | 39.5 |

## Agent readiness — 7.1 (agent-aware)

| Dimension | Value |
|---|---|
| Spec Presence | no |
| Agentic Access | no |
| Reversibility Documented | no |
| MCP Server | documented |
| Auth Clarity | no |
| Idempotency | no |
| Error Semantics | no |
| OpenAPI Examples | no |
| Rate Limit Signal | documented |
| Event Surface Described | no |
| Agent Skills | no |
| Well Known Catalog | no |
| Consent Identity | no |
| Agent Card | no |
| Dry Run Mode | no |

## Access

Free — onboarding: unknown, pricing: free, trial: no (confidence: medium).

## APIs (21)

- **Kong AI Gateway** — Kong AI Gateway is the connectivity and governance layer for AI-native applications. Built on Kong Gateway, it provides a universal LLM API across 16+ providers, semantic cachin...
- **AI Proxy Plugin** — The AI Proxy plugin transforms and proxies requests to a configured AI provider and model, shielding client applications from provider-specific request and response shapes.
- **AI Proxy Advanced Plugin** — The AI Proxy Advanced plugin extends AI Proxy with load balancing, weighted distribution, and fallback across multiple providers and models simultaneously.
- **AI Rate Limiting Advanced Plugin** — Token-aware rate limiting tailored for LLM traffic, with per-consumer and per-model budgets rather than just request counts.
- **AI Prompt Guard Plugin** — Enforces allow- and deny-lists for prompts and text completions, blocking disallowed content before it reaches the model.
- **AI Semantic Prompt Guard Plugin** — Topic-aware variant of AI Prompt Guard that classifies prompts by meaning and blocks restricted topics regardless of phrasing.
- **AI PII Sanitizer Plugin** — Detects and redacts personally identifiable information from prompts and responses traversing the gateway.
- **AI Semantic Cache Plugin** — Caches LLM responses by prompt similarity so semantically equivalent requests can be served from cache, reducing latency and provider spend.
- **AI RAG Injector Plugin** — Automates retrieval-augmented generation by injecting retrieved context into prompts at the gateway, so application code does not need to implement RAG plumbing.
- **AI Prompt Template Plugin** — Provides reusable, fill-in-the-blank prompt templates managed at the gateway layer.
- **AI Prompt Decorator Plugin** — Prepends or appends messages to chat history before requests reach the model.
- **AI Prompt Compressor Plugin** — Reduces prompt token count before forwarding to the provider to lower latency and cost.
- **AI Azure Content Safety Plugin** — Integrates Azure AI Content Safety for content moderation on prompts and responses.
- **AI AWS Guardrails Plugin** — Integrates Amazon Bedrock Guardrails for safety enforcement on traffic passing through Kong AI Gateway.
- **AI GCP Model Armor Plugin** — Integrates Google Cloud Model Armor for safety inspection on prompts and responses.
- **AI Lakera Guard Plugin** — Integrates Lakera Guard for prompt-injection and jailbreak detection.
- **AI Semantic Response Guard Plugin** — Inspects model responses by meaning and blocks responses that violate configured semantic policies.
- **AI Custom Guardrail Plugin** — Lets operators define custom guardrail logic for prompts and responses without writing a full Kong plugin.
- **AI Request/Response Transformer Plugin** — Uses LLMs at the gateway to transform request and response payloads (for example, reshaping JSON or translating fields) on the data path.
- **Kong Agent Gateway** — Kong Agent Gateway is a capability of Kong AI Gateway (GA April 2026 with AI Gateway 3.14) that governs agent-to-agent (A2A) communication. It enforces agent identity verificati...
- **Kong MCP Registry** — Kong MCP Registry (launched February 2026) is an enterprise directory inside Kong Konnect for registering, discovering, and governing MCP servers and AI-native tools. It provide...

## MCP servers (1)

- **Kong Konnect MCP Server**

## Security (1)

- **Kong Ai Gateway Domain Security** — TLSv1.3 · HSTS · DNSSEC · DMARC

## Plans (1)

- **Kong Ai Gateway Plans Pricing**

## Use cases (5)

- **Provider-Agnostic LLM Access** — Give applications a stable LLM endpoint while swapping providers and models behind the gateway.
- **AI Cost Control** — Apply token budgets, semantic caching, and prompt compression to keep LLM spend bounded.
- **AI Safety and Compliance** — Enforce prompt firewalls, PII redaction, jailbreak detection, and content safety on every prompt and response.
- **Agentic Tool Governance** — Govern which MCP tools agents can discover and call, and inspect agent-to-agent traffic in production.
- **RAG at the Edge** — Inject retrieval context into prompts at the gateway without changing client code.

## Tags

AI Gateway, LLM, MCP, A2A, AI Governance, Konnect, Kong

---

Profiled by [API Evangelist](https://apievangelist.com) and published on [APIs.io](https://apis.io/providers/kong-ai-gateway/). Scores are computed from the provider's own public artifacts under a published rubric.