Nutanix Agent Gateway
by Nutanix
A single control point for what your AI agents may call, and what it costs
Nutanix Agent Gateway is a governance and cost-control plane that sits between AI agents, the models they call and the Model Context Protocol servers they reach enterprise tools through. It gives platform teams one place to enforce access policy, meter token spend and audit every request, across both frontier models in public cloud and privately hosted models on their own Kubernetes.
Nutanix Agent Gateway is the agent governance layer of Nutanix Enterprise AI (NAI), generally available since NAI 2.7 on 26 May 2026 and extended in NAI 2.8 in August 2026. It is a centralised gateway that manages and secures two directions of traffic: from agents out to large language models, and from Model Context Protocol servers in to the business tools and private data sources those agents want to touch. Its Unified API exposes external provider models and self-hosted models behind one inference endpoint, so an organisation can route a document summarisation to a cheap open-weight model while reserving a frontier model for complex multi-agent work — the specific waste Nutanix CEO Rajiv Ramaswami described as a 'free-for-all' in which anybody can call anything. The MCP Gateway added in 2.8 applies granular access control over MCP servers, with fine-grained identity and access management, custom roles, model-sharing controls and support for air-gapped NVIDIA NIM deployments. Unified Observability centralises visibility into token usage, MCP server access and LLM activity; audit logs record every MCP request; and granular token-based rate limiting enforces quotas centrally rather than per team. The gateway runs on any CNCF-certified Kubernetes — Nutanix Kubernetes Platform, AWS EKS, Azure AKS, Google GKE, bare metal or a dark site — and does not require Nutanix HCI. It was announced as part of Nutanix Agentic AI at NVIDIA GTC on 16 March 2026 alongside Nutanix Private Inference, which 2.8 extended with parameter-efficient fine-tuning, multi-GPU and batch inference and speculative decoding. Hardware support is NVIDIA-first today, with AMD support targeted for the end of 2026.
The platform or infrastructure team at a Nutanix-adjacent enterprise that already runs Kubernetes and is watching agent token spend and MCP access sprawl outrun its governance.
One control point that enforces which agents may call which models and tools, meters token consumption against quotas, and produces an audit trail of every MCP request.
At a Glance
- Category
- AI Agents & Orchestration
- Pricing
- Subscription, Usage-based, Contact for pricing
- Target Market
- CIOs, CTOs, Platform Engineers, Infrastructure Architects, AI Platform Teams
- Deployment
- Hybrid, Self-hosted, Multi-cloud, Edge-first
- Founded
- 2009
- Headquarters
- San Jose, California, United States
- Team Size
- 500+
- Customers
- Nutanix reports more than 29,000 customers across its platform; Agent Gateway adoption is not separately disclosed
Key Features
- ✓Unified API for models
Exposes external provider models and self-hosted models through one endpoint so teams can pick the right model per use case without rewriting applications.
- ✓MCP Gateway governance
Applies granular access control to Model Context Protocol servers, so an agent reaches only the business tools and data sources its role permits.
- ✓Granular token-based rate limiting
Enforces token quotas and limits centrally, which is what converts agent spend from an unbounded surprise into a budgeted line item.
- ✓Unified observability
Centralises visibility into token usage, MCP server access and LLM activity across cloud-hosted and private models in one view.
- ✓Audit logging
Records every MCP request with a comprehensive audit trail, which is the evidence an AI governance or compliance review actually asks for.
- ✓Fine-grained IAM with custom roles
Added in NAI 2.8, it lets platform teams define custom roles and model-sharing controls rather than relying on coarse project-level access.
- ✓Air-gapped and private inference support
Runs with self-hosted models and air-gapped NVIDIA NIM deployments, so regulated and sovereign workloads never leave the customer's boundary.
Capabilities
Use Cases
- •Stopping agent token spend running away
Route trivial summarisation to a cheap model and cap each team's token quota centrally, instead of discovering the overspend on the next cloud bill.
- •Governing which tools an agent may touch
Use MCP access control so a support agent reaches the ticketing system but never the finance database, with every request logged.
- •Running agents in an air-gapped environment
Serve private models and NVIDIA NIM microservices inside a dark site so regulated data never leaves the customer's own infrastructure.
- •Consolidating multi-provider model access
Put frontier cloud models and self-hosted open-weight models behind one API so applications are not rewritten when the model choice changes.
- •Evidencing AI governance to auditors
Export audit logs of agent, model and MCP activity when an internal risk function asks how autonomous systems are controlled.
Ideal For
Best For
- ✓Centralising access control over which agents may reach which MCP servers and enterprise data sources
- ✓Enforcing token quotas and rate limits so agent experiments cannot run up an unbounded inference bill
- ✓Routing cheap tasks to open-weight models and expensive tasks to frontier models behind one unified API
- ✓Producing an auditable record of every agent, model and MCP tool call for AI governance and compliance review
- ✓Running agent infrastructure in air-gapped or sovereign environments using self-hosted models and NVIDIA NIM
Not Ideal For
- ✗Small teams without a Kubernetes platform practice — this is infrastructure software that assumes a CNCF-certified cluster and people to run it, not a SaaS sign-up
- ✗Buyers who need to size cost before talking to a vendor; no list pricing is published and NAI is metered by aggregate GPU RAM or vCPU, which is hard to model from outside
- ✗Organisations standardised on AMD or Google TPU accelerators today, since the gateway is NVIDIA-based with AMD support only targeted for end of 2026
- ✗Teams that want to evaluate for free before committing — Nutanix confirms there is no free trial, only a hosted proof-of-concept via its Test Drive programme
Deployment
Market & Ratings
Nutanix reports more than 29,000 customers across its platform; Agent Gateway adoption is not separately disclosed
Market Analysis
Pros
- ✓Addresses the two problems platform teams actually raise about agents — uncontrolled model spend and ungoverned tool access — in one component rather than two products
- ✓Infrastructure-agnostic: any CNCF-certified Kubernetes, public cloud, bare metal or air-gapped, with no Nutanix HCI requirement
- ✓Shipped fast and kept shipping — GA in NAI 2.7 in May 2026, with MCP Gateway, custom IAM roles and air-gapped NIM support added in 2.8 by August 2026
- ✓Backed by a public company with 7,800 employees and 29,000+ customers, and validated with Cisco, Dell, Lenovo, Supermicro and Fujitsu hardware
Cons
- ✗Hardware support is NVIDIA-only at launch; AMD support was still a year-end 2026 target and TPU support only an aspiration, so accelerator choice is constrained today
- ✗No published pricing and no free trial — evaluation runs through a sales-led proof-of-concept, which is slow compared with the API gateways it competes against
- ✗GPU-RAM-based metering is difficult to forecast before deployment, and Nutanix licensing complexity is a recurring complaint in reviews of the company's other products
- ✗Agent Gateway is too new to have independent user reviews; no dedicated G2, PeerSpot or TrustRadius entry existed for it as of September 2026, so all capability claims are vendor-sourced
- ✗It is a control plane, not a security product — it does not detect prompt injection or model-level attacks, so it sits alongside an AI security layer rather than replacing one
Pricing
NAI Standalone
Contact for pricing
- ✓Runs on any CNCF-certified Kubernetes
- ✓Agent Gateway and MCP Gateway
- ✓Metered by aggregate GPU RAM in 1 GB increments, or by vCPU where no GPU is present
NAI and NKP (NAI-NKP)
Contact for pricing
- ✓Enterprise AI Pro bundled with Nutanix Kubernetes Platform
- ✓Available in Pro or Ultimate tiers
NAI Full Stack (NAI-GPT)
Contact for pricing
- ✓Validated stack including NAI Pro, NCI, NKP, NDK, NUS and NDB
- ✓For customers building a complete on-premise AI factory
No list pricing is published anywhere. Nutanix Enterprise AI, which is how Agent Gateway is licensed, is metered by aggregate GPU RAM for inference clusters with GPUs installed — sold in 1 GB entitlement increments — or by the vCPUs of worker nodes where no GPU accelerator is present. It comes in three packages: NAI Standalone, NAI plus NKP, and NAI Full Stack. There is no free trial; Nutanix offers a hosted proof-of-concept through Test Drive instead. NVIDIA AI Enterprise licensing may be a separate per-GPU cost on top.
Security & Compliance
Sources
This page was written from 9 sources, 4 on domains other than nutanix.com.
- 1.nutanix.com — introducing nutanix agent gatewayvendor
- 2.nutanix.com — nutanix enterprise aivendor
- 3.nutanix.com — nutanix unveils nutanix agentic aivendor
- 4.nutanix.com — software optionsvendor
- 5.nutanix.com — faqvendor
- 6.sdxcentral.com — nutanix targets enterprise ai wallet bleed with agent contro
- 7.technode.global — nutanix enterprise ai 2 8 mcp gateway agentic ai
- 8.itbrief.com.au — nutanix launches agent gateway to govern ai agents
- 9.en.wikipedia.org — Nutanix
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Maven AGI
Enterprise AI agents for customer support, sales and retention that act across your helpdesk, CRM and phone systems
Yellow.ai
Agentic AI platform for voice, chat and email customer service, now extended to employee IT and HR support with Nexus EDGE
Siena AI
AI customer service, shopping and social agents for consumer brands, sharing one memory of every order and conversation
Hermes Business
Hosted Hermes Agent for whole teams: one shared balance, per-member spend caps and a shared skills library