Kong AI Cost Management and AI Registry: What they mean for production AI

The industry is hitting an inflection point. AI adoption has outpaced the tooling to govern it. Organizations are flying blind on spend, and agents are being hand-wired to tools in ways that don't scale.

At Kong's API & AI Summit 2026, two announcements stand out as the most immediately impactful for enterprises running AI in production: AI Cost Management and AI Registry.

Together, they address the two hardest problems in operationalizing agentic AI: can we afford it, and can we safely let agents discover what they need?

Problem 1: The AI attribution gap is bleeding budgets

The scenario is familiar: Your engineering team burns through the annual AI budget by April. Finance sees a $500K invoice from one model provider and another $300K from a second. Nobody can explain which application, team or agent drove that spend, much less whether it delivered return on investment (ROI).

The root cause: Provider dashboards show tokens and models. Businesses operate in teams, applications, cost centers and agents.

This is the attribution gap, and it's where Kong's AI Cost Management (announced in Early Access at the summit) is aimed.

What Kong AI Cost Management actually does

Kong sits at a unique architectural position: directly between AI consumers and the models they call. That means the gateway sees every interaction: the consumer identity, the application, the model, the provider, the token counts and the cost in real time.

Kong builds a five-layer economic model on top of that visibility:

The strategic portability win

For multi-provider reality, this is what matters: your business context stays intact when you swap models.

The Customer Experience team is still the Customer Experience team whether they're calling one vendor's frontier model today, a competitor's model tomorrow or a self-hosted open-weight model on-prem next quarter. Kong separates the business attribution model from the AI infrastructure layer.

Change the model. Change the provider. Change where it runs. The cost center, the team, the application, all of it remains queryable and governable.

Routing to the right model at the right time

This is where the "business logic" piece comes in. Kong AI Gateway 2.2, which reached general availability (GA) the same day, adds:

• Intelligent LLM routing: route based on cost, latency, quality, task type
• Credential-based rate limiting: granular controls per API key, per team, per application
• Passthrough mode: govern custom/self-hosted models (vLLM, Ollama, NVIDIA NIM™) without waiting for native support
• TypeSafe AI Jev support: Jev is TypeSafe AI's decision model. It returns typed, probabilistic decisions that software can act on directly, rather than free-form text. Gateway 2.2 adds native support for TypeSafe Jev as a provider and format, including its decisions capability, which can inform intelligent model routing.
• NVIDIA NeMo™ Switchyard integration: NeMo™ Switchyard is an open-source model-selection library from NVIDIA. It runs as a separate selection service rather than inside the gateway. Kong AI Gateway supplies the request context, Switchyard's classifier returns a model choice based on the intent and complexity of the request, and the gateway routes accordingly — a small local model for simple queries, a specialized model for code generation or a frontier model for complex reasoning.
• Headroom integration (Tech Preview): prompt compression that reduces token consumption before requests reach the model

Result: A query about "reset my password" routes to a fast, cheap model (potentially a small local open-weight model served via NVIDIA NIM™). A complex legal contract analysis routes to your best reasoning model. A code generation task routes to a specialized coding model. The gateway decides — informed by Switchyard's semantic understanding — and the developer doesn't hardcode it.

Problem 2: Agents are being hand-wired into a corner

The other side of the production coin: You have 50 agents in development. Each needs models, Model Context Protocol (MCP) servers, tools and other agents. Today, developers manually configure every connection.

That doesn't scale. Gartner predicts that by 2028 an average global Fortune 500 enterprise will have over 150,000 agents in use, up from fewer than 15 in 2025, and that only 13% of organizations think they have the right AI agent governance in place. Manual wiring is already broken.

AI Registry: Self-service discovery for agents, with governance built in

Kong AI Registry is the Developer Portal equivalent for agents.

The lifecycle: Register → Approve → Publish → Discover → Consume

What gets registered: Models, MCP servers, agents, tools, resources and skills, each with ownership, version, approval status, risk classification, authentication, policies and access controls.

Governed views, not one giant directory:

• Engineering agents discover coding models, GitHub MCP servers, code-review skills
• Finance agents see approved financial models and ERP tools
• Contractors get a narrow, explicitly approved subset
• Partners see only what's published for external consumption

The critical distinction: Registry ≠ Catalog

Kong also announced Konnect Catalog, now GA: the human-facing system of record for everything (APIs, models, MCP servers, agents and event streams).

An asset in Catalog doesn't automatically become available to agents. Publishing it through a Registry is what makes it discoverable, and only to the audiences you approve.

This is least-privilege access for agents, enforced at the discovery layer.

The multiplier: Kong Context Management + Registry = safe MCP at scale

Neither capability reaches its ceiling alone. The third piece is Kong Context Management, which exposes existing APIs as agent-consumable tools without exploding tool count or rebuilding the permissions model.

Code Mode (GA as part of Kong Context Management) exposes APIs as MCP interfaces without exploding the tool count:

• Traditional: API (200 ops) → 200 MCP tools → 200 tool definitions in every agent context
• Code Mode:  API → compact MCP interface → agent searches → retrieves schema on demand → executes

The access control carry-through: Your existing REST API already enforces authentication, authorization, rate limits and audit logging. Since Kong Context Management generates the MCP interface on top of that API, requests via MCP pass through the same enforcement layer. You don't rebuild permissions. You don't duplicate policy.

Combine with Registry: The MCP server gets registered, approved, published to the right agent audience. The agent discovers it at runtime. The gateway enforces the same policies whether the call comes via REST or MCP.

The combined effect is that organizations can safely expose hundreds of existing APIs as agent tools without the context-window tax, the token-cost explosion or the access-policy reimplementation nightmare.

Why these two capabilities belong together

Cost Management gives you the feedback loop: spend → attribution → business value → optimization → repeat.

Registry gives you the control plane: register once → govern centrally → agents discover safely → no manual wiring.

Both run on the same platform. Same gateway. Same policies. Same observability. Same identity layer.

What this means for your organization

If you're a platform team

• Custom cost dashboards become unnecessary; the gateway does attribution natively.
• Manual agent wiring gives way to governed self-service through Registry and Catalog.
• Access policies for MCP don't need to be reimplemented. Kong Context Management exposes your existing APIs through MCP, so the underlying API enforcement layer carries through.

If you're a financial operations (FinOps) or finance leader

• Finally answer: "What did we get for our AI spend?", down to cost per business outcome (cost per case resolved, cost per code review, cost per fraud investigation).
• Budget by team/application, not by provider invoice.

If you're a security or governance lead

• Least-privilege for agents via Registry's scoped views
• Credential-less access via Token Vault (coming soon): agents access sensitive systems without holding long-lived secrets
• Audit trail across every model call, tool invocation, agent interaction

If you're building agents

• Discover, don't configure. Query the Registry at runtime for available models, tools, MCP servers.
• Compose via code. Kong Context Management lets agents execute multi-step workflows (loops, conditionals, error handling) without round-tripping to the model for every tool call.
• Trust the routing. The gateway picks the right model for the task, you focus on the agent logic.

The bigger picture: governance now includes economics

Kong's framing at the summit was deliberate: "As AI becomes a fundamental part of enterprise infrastructure, governance can no longer mean security alone."

Governance now means:

1. Who/what can access AI (Identity, Token Vault, Registry)
2. What AI systems are doing (Observability, Catalog, Registry)
3. How models, APIs and tools are consumed (Policies, Rate limiting, Routing)
4. The economics of it all (Cost Management: price, attribute, plan, control, optimize)

The gateway becomes the unified control  point for all four.

What's available now vs. coming

The bottom line

AI Cost Management solves the "we're spending too much and don't know why" problem, with multi-provider attribution, business-context budgeting, inline controls, and measured optimization.

AI Registry solves the "agents can't scale with manual wiring" problem, with governed self-service discovery, least-privilege access, and a clean separation between what exists (Catalog) and what this agent may use (Registry).

Together, they're the foundation for running AI agents in production at enterprise scale.

Not "someday." The Catalog and Kong Context Management are GA today. The Cost Management and Observability early access programs are open now. The Registry is in limited release, with full GA expected imminently. If budget accountability is your most pressing blocker, the Cost Management early access program is your fastest entry point, and you can get hands-on with Kong AI Gateway today in the AI Gateway hands-on lab with Kong.

Further reading (all Summit-day announcements)

• API & AI Summit 2026 Launch Recap: Master overview of all 6 stages
• Take Control of the Economics of AI:  AI Cost Management deep dive
• Introducing Kong AI Registry: Registry architecture and FAQ
• Introducing Kong Context Management: Code Mode and how agents compose multi-step API workflows
• Introducing Kong Catalog: Platform-wide inventory
• Introducing Advanced AI Observability: Session-level debugging
• Kong AI Gateway 2.2: Strategic portability, Jev, Skills, passthrough, custom plugins, NVIDIA NeMo™ Switchyard integration