Programmable AI Control Plane Vendors for CI/CD Pipelines
The short answer
A control plane belongs in CI/CD only if every route, provider target and policy can be written as a file, diffed on a pull request and applied on merge. Five vendors do that well as of November 16, 2026:
- Kong AI Gateway — best if you already run Kong. decK (v1.70.1, released November 12, 2026) manages services, routes and AI plugins as YAML with
deck validate,deck diffanddeck sync. - agentgateway — best for Kubernetes platform teams. Apache-2.0 data plane (v1.7.0, October 27, 2026) built on the Kubernetes Gateway API; LLM, MCP and A2A routing are CRDs that Argo CD or Flux can apply.
- Portkey — best SaaS gateway with Terraform. Its provider manages configs (routing, fallbacks, load balancing), guardrails, usage and rate-limit policies, API keys and secret references.
- LiteLLM — best open-source proxy for many providers. One
config.yamlfor models, credentials and routing; also a Management API and Terraform. - Cloudflare AI Gateway — best edge option. The
cloudflare_ai_gatewayTerraform resource sets caching, rate limits, logging, DLP and guardrails.
This page is the CI/CD spoke of our guide to trusted agent control plane providers, which covers enterprise suites such as Microsoft Agent 365 and AWS AgentCore.
The comparison
| Vendor | What you write | How the pipeline applies it | Licence / price (Nov 16, 2026) | Best for |
|---|---|---|---|---|
| Kong AI Gateway | decK YAML: services, routes, plugins (AI Proxy, AI Proxy Advanced, prompt guards, MCP and A2A plugins) | deck validate → deck diff on PR → deck sync on merge | decK Apache-2.0; AI Proxy Advanced only in AI Gateway Enterprise (quote) | Teams already standardised on Kong |
| agentgateway | Gateway API Gateway/HTTPRoute plus AgentgatewayPolicy, AgentgatewayBackend (agentgateway.dev/v1alpha1) | kubectl apply or a GitOps controller | Apache-2.0, free; enterprise support from Solo.io | Kubernetes-first platform teams |
| Portkey | Terraform: configs, guardrails, prompts, usage and rate-limit policies, secret references, SCIM mappings | terraform plan on PR, terraform apply on merge | Developer free (10k logs/month); Production $49/month; Enterprise custom | SaaS buyers who govern many teams |
| LiteLLM | config.yaml (model_list, router_settings); Management API; Terraform | Redeploy proxy with new config, or Terraform | MIT outside the enterprise/ directory | Self-hosted multi-provider routing |
| Cloudflare AI Gateway | Terraform cloudflare_ai_gateway: cache TTL, rate limiting, log collection, DLP, guardrails | terraform apply | Part of a Cloudflare account | Apps already on Cloudflare |
Vendor notes
Kong — extend the gateway you already automate
Kong’s GitOps pattern is documented by Kong itself: a pull request runs deck validate and deck diff, so reviewers see exactly which routes and plugins change, and the merge runs deck sync. AI traffic uses the same objects as API traffic, so an AI route is one more service with an AI Proxy plugin attached. The catch is licensing: AI Proxy Advanced — load balancing across models and semantic routing — is available only in Kong’s AI Gateway Enterprise offering, so price it before you design around it.
agentgateway — the Kubernetes-native option
agentgateway is an Apache-2.0 data plane designed for agents, MCP tools and LLMs, and it implements the Kubernetes Gateway API rather than a proprietary config format. Routes are standard HTTPRoute objects; AI-specific behaviour (prompt guards, retries evaluated with CEL expressions, backends for model providers and A2A agents) lives in AgentgatewayPolicy and AgentgatewayBackend resources. That makes it the cleanest GitOps fit of the five: the same Argo CD or Flux pipeline that ships your services ships your agent routes. The API version is still v1alpha1, so expect field changes between releases.
Portkey — Terraform for a SaaS gateway
Portkey’s official Terraform provider talks to its Admin API and covers the objects platform teams want under review: workspaces and members, provider integrations and virtual keys, gateway configs with routing, fallbacks and load balancing, guardrails, prompt templates, usage-limit and rate-limit policies, and secret references to AWS Secrets Manager, Azure Key Vault or HashiCorp Vault. Pricing as of November 16, 2026: Developer is free with 10,000 recorded logs a month (not for production), Production is $49 a month with 100,000 logs and $9 per extra 100,000, and Enterprise is custom.
LiteLLM — one file for many providers
LiteLLM’s proxy reads a single config.yaml that defines models, provider credentials (from environment variables), routing strategy, logging and server settings. LiteLLM documents four ways to manage it — YAML, the Admin UI, the Management API and Terraform — so CI can either rebuild the proxy with a new file or call the API. The code outside the litellm/enterprise/ directory is MIT; enterprise features carry their own licence.
Cloudflare AI Gateway — config at the edge
The cloudflare_ai_gateway Terraform resource creates a gateway per account and sets cache behaviour, log collection, rate limiting, billing mode, DLP rules and guardrail categories. It is the simplest of the five, and the right one when your application and its secrets already live on Cloudflare.
How to choose
- Already run Kong for APIs? Stay on Kong and budget for AI Gateway Enterprise if you need multi-model balancing.
- Kubernetes and GitOps everywhere? agentgateway.
- Buying SaaS, governing many teams, Terraform in place? Portkey; Cloudflare if you are already on Cloudflare.
- Want open source in front of 100+ model providers? LiteLLM.
Whichever you choose, enforce three rules in the pipeline: no API keys in the config files (use secret references), a required diff review on every policy change, and a canary route before shifting traffic to a new model. For background, see managed vs self-hosted agent gateways and the best LLM gateways.
Last verified: November 16, 2026.