New VDF AI Enterprise is now available on AWS Marketplace
Enterprise AI Comparison

Azure AI Foundry Alternative for
Agents That Run Anywhere

Microsoft Foundry, still searched as Azure AI Foundry, runs agents as a managed service in Azure regions and binds them to Entra, Azure Policy and Azure billing. VDF AI runs governed agents in its own cloud, your Azure tenant, another cloud, your data center or an air-gapped enclave. This page sets out which workload runs where, checked against Microsoft’s documentation in September 2026.

Book a Demo See the Verdict
QUICK VERDICT

The Short Answer

Microsoft Foundry is the better fit when your applications, data and identity already live in Azure and Microsoft 365, your developers want Microsoft to host prompt agents or containerised agent code, and your regulators accept agents executing in an Azure region under Entra identities and Azure Policy.

VDF AI is the better fit when some agents must run on infrastructure you operate, sometimes with no connection at all, when models come from several providers or from your own GPUs, and when governance and cost ceilings have to survive a future move away from any one cloud.

Microsoft Foundry
VDF AI
Best for
Azure-native agent engineering
Governed agents on any infrastructure
Where agents execute
Microsoft-managed compute in Azure regions
VDF Cloud, your tenant, your data center, air-gapped
On-prem agent layer
Agentic Retrieval in Foundry Local (preview)
Full platform as self-hosted packages
Model catalogue
10,000+ catalog models
Approved providers plus your own open weights
Agent identity
Dedicated Entra identity per hosted agent
Per-role tool grants; SSO on Enterprise Cloud; Entra ID SSO on-premise
Pricing mechanic
Tokens, tool meters, sandbox compute
Per user (Cloud) or annual capacity (on-prem)
Cost of leaving
Identity, tools and traces are Azure resources
Same product on any Kubernetes
PRICING & DEPLOYMENT

Foundry Metering and Deployment Scope

Microsoft bills each Foundry component on its own meter; VDF AI sells access by user or by an agreed capacity.

Microsoft Foundry Pricing

Checked 27 Sep 2026 on the Foundry pricing page and the Agent Service pricing page

PlatformFree to exploreEach feature you use bills at its own rate
ModelsPer tokenRate depends on model and deployment type
Prompt agentsNo agent feeTokens and tool calls still charged
Hosted agentsvCPU-hour + GiB-hourMeasured across active session sandboxes
Built-in toolsMeteredVector storage per day, interpreter sessions, search calls

Hosted-agent compute follows sessions, and Microsoft’s own sizing notes warn that an oversized sandbox multiplies cost by concurrency. Plan tokens, tools, tracing ingestion and storage as separate budget lines. Microsoft also lists a one-year Agent pre-purchase plan on the Foundry pricing page.

VDF AI Pricing

Mechanics are published on the VDF AI pricing page

CloudPer userFree Starter, Professional, Enterprise Cloud
On-premisesAnnual capacityOne pool per organisation, unlimited people
Two metersTokens and platform transactions; the bundle binds on whichever is reached first
Never meteredRouting, policy checks, audit logging, index retrieval
ServicesImplementation is scoped and quoted apart from the licence

The on-premises figure is fixed before the term starts. A longer agent run draws more tokens from the pool, yet the governance work wrapped around it never adds to the count.

What runs where, stated precisely

Foundry Agent Service is an Azure platform service: prompt agents and hosted agents execute on Microsoft-managed compute in Azure regions, and virtual-network isolation plus bring-your-own storage, Azure AI Search and Cosmos DB are the controls for keeping data close. Microsoft’s on-premises story is a different product line. Foundry Local on Azure Local serves models on Arc-enabled Kubernetes, and Agentic Retrieval in Foundry Local adds local RAG agents with an MCP server; both were in preview when we checked. VDF AI ships the same product surface to its cloud and to self-hosted clusters, which is why the on-premises deployment is not a reduced edition.

GOVERNANCE

Identity, Policy & Evidence

Both platforms take governance seriously; they anchor it to different things.

Identity
FoundryA dedicated Microsoft Entra identity per hosted agent, Azure RBAC, and On-Behalf-Of passthrough for delegated calls
VDF AINo per-agent directory identity; users sign in with Entra ID on-premise or with SSO on Enterprise Cloud, and access is governed at the tool
Tool access
FoundryToolboxes publish curated tools behind one managed MCP endpoint with central authentication and versioning
VDF AIAn MCP gateway registry where administrators grant tools to roles and every call is audited
Runtime guardrails
FoundryContent filters, cross-prompt injection mitigation and Azure Policy across Foundry resources
VDF AIAllowed-model lists, allowed external services, per-run and monthly budget caps, and Human Approval nodes
Data location
FoundryChoice of region and deployment type decides where inference happens; BYO storage keeps state in your resources
VDF AISelf-hosted inside your perimeter; with local models nothing has to leave the network at all
Audit evidence
FoundryEnd-to-end tracing, evaluations and Application Insights dashboards for every model and tool call
VDF AIA durable Vault record per execution, from which an EU AI Act or sector evidence pack assembles
Spend control
FoundryModel quota per region in the Foundry control plane, plus an optional pre-purchase plan
VDF AIHard budget caps per network, inside a capacity pool agreed for the year

The wider control model is described on the AI governance platform page.

BUILDING AGENTS

Foundry Agent Service vs VDF AI Agents

Foundry is built for developers who want Azure to host their agent code; VDF AI is built for teams that want a governed product around the agent.

How Foundry approaches it

  • Three agent types — prompt agents defined by configuration, voice-based prompt agents (preview), and hosted agents that run your own container
  • Framework choice — hosted agents accept Microsoft Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic Agent SDK or custom code
  • Toolboxes — web search, file search, code interpreter, MCP servers and OpenAPI tools curated once and shared
  • Distribution — publish to Teams, Microsoft 365 Copilot and the Entra Agent Registry; A2A v1.0 is generally available
  • Hosted in Azure — hosted-agent sandboxes run only in the Azure regions Microsoft lists for them
  • Government cloud gap — Microsoft’s region guide lists Azure AI Agents as unsupported in US Gov regions (page dated August 2026)

How VDF AI approaches it

  • VDF AI Agents — agents assembled from reusable Agent Skills, governed tools and connected knowledge
  • VDF AI Networks — staged multi-agent work with visible intermediate outputs, schedules, webhooks and Human Approval nodes
  • Smart routing — Auto, Sustainable and Regulated modes choose a model per step from the list an admin permits
  • Microsoft reach — one Microsoft sign-in covers OneDrive, Outlook, SharePoint, Planner, Azure DevOps and Dynamics 365
  • Beyond Microsoft — Jira, Confluence, GitHub, Slack, Notion and Google connectors under the same audit trail
  • Smaller catalogue — VDF AI routes between the providers and local models you approve rather than fronting a 10,000-model marketplace

Weighing Microsoft’s low-code builder rather than Foundry? Read the Copilot Studio comparison, or the broader guide to Copilot alternatives.

ARCHITECTURE

Where Each Platform Can Run

The deciding question for regulated buyers is rarely features. It is location.

Microsoft Foundry

Azure PaaS with preview edge options

  • Foundry projects — created in listed Azure regions, plus US Gov Arizona and Virginia
  • Agent Service — prompt and hosted agents on Microsoft-managed compute
  • Private networking — VNet isolation, a BYO VNet for hosted agents, BYO storage and Cosmos DB
  • Foundry Local on Azure Local — model serving on Arc-enabled Kubernetes (preview, by request)
  • Agentic Retrieval in Foundry Local — local RAG agents, knowledge bases and an MCP server (preview)

Azure Local disconnected operations were announced as available worldwide in February 2026. Microsoft documents the local agent layer as its own Arc extension with a narrower feature set, not as Agent Service running on your hardware.

VDF AI

One product, four landing zones

  • VDF Cloud — the managed service, operated by VDF AI
  • Self-hosted — signed container packages from portal.vdf.ai on Docker Compose or any conformant Kubernetes, AKS included
  • Private or sovereign cloud — the same packages inside a tenancy you control
  • Air-gapped — images mirrored into your internal registry, with the router’s air-gap mode keeping calls local
  • State — PostgreSQL 16+, self-run or managed, Azure Database for PostgreSQL among the named options

Changing the landing zone changes who runs the servers, not which features the agents get. Sizing guidance lives in the on-prem reference architecture.

DEPLOYMENT

Deployment Scenarios

Seven places an enterprise may need an agent to run, and what each vendor documents for them.

ScenarioMicrosoft FoundryVDF AI
Azure commercial regionFoundry Agent Service, prompt and hosted agentsSelf-hosted on AKS or VMs in your subscription
Azure Government (US)Foundry projects exist; Azure AI Agents listed as unsupportedSelf-hosted on compute you already operate
Azure Local, connectedFoundry Local serving and Agentic Retrieval, both previewSelf-hosted on any conformant Kubernetes cluster
Fully disconnected siteAgentic Retrieval under the disconnected-operations previewAir-gapped install from your internal registry
Another public cloudAgents still execute in Azure; callers connect over the networkSame packages; AWS Marketplace listing for the self-hosted edition
Data center without Azure LocalNo managed runtime; open-source Agent Framework is a do-it-yourself pathSelf-hosted platform with vendor support
Who runs upgradesMicrosoft, as a managed serviceVDF AI on Cloud; your team on its own cadence when self-hosted
FAIR PLAY

When to Choose Microsoft Foundry

For an Azure-first organisation, Foundry is a strong default and often the right one.

Foundry is the right call when…

  • Your data, apps and identity already sit in Azure and Entra, and an Azure region satisfies your regulators.
  • You would rather not operate Kubernetes for agents and want scale-to-zero sandboxes managed for you.
  • Your developers already write agents with Agent Framework, LangGraph or the OpenAI Agents SDK and want Microsoft to host them.
  • You want the widest model catalogue under a single Azure agreement, first-party and partner models side by side.
  • Agents must appear in Teams and Microsoft 365 Copilot with the least integration work.
  • Voice agents with telephony channels belong in the same platform, and you can live with their preview status.
Where Foundry is genuinely strong
Per-agent Entra identity

Each hosted agent is issued its own identity at deploy time, so downstream access is scoped and attributable without shared secrets.

Catalogue breadth

More than 10,000 models from Microsoft, OpenAI, Anthropic, Meta and others sit behind one project endpoint.

Managed home for your code

Per-session VM-isolated sandboxes persist files between turns, scale automatically and emit OpenTelemetry traces.

Microsoft 365 distribution

Published agents reach Teams and Copilot, and Copilot Studio can call a Foundry agent through a preview connection.

DECISION SIGNALS

When VDF AI Fits Better

Signals that a workload needs something other than an Azure-hosted agent runtime.

The site never reaches Azure

A plant floor, a ministry network or a classified enclave may have no path to a cloud region. VDF AI installs from your own registry and keeps working with no outbound connection.

Government cloud is mandatory

If agents must live in US Gov regions, check Microsoft’s feature list first: in August 2026 it marked Azure AI Agents as unsupported there. Self-hosted VDF AI runs on compute you already operate.

Models must run on your GPUs

Open-weight models served by vLLM, Ollama or llama.cpp sit next to Azure OpenAI in one router, with a per-network list of what each workflow may call.

Finance wants one number

Tokens, tool meters and sandbox hours accumulate per session. An annual capacity pool gives finance a figure before the year begins, with routing, policy and audit outside the meter.

Exit risk is on the risk register

Foundry agents lean on Entra agent identities, Toolboxes, Foundry conversation storage and Application Insights. VDF AI behaves the same on AKS, another cloud or bare metal.

Workflows cross ecosystems

One flow that reads Confluence, updates Jira, posts in Slack and files to SharePoint gets a single audit trail when every hop passes through VDF AI connectors and its MCP gateway.

For how Microsoft, Google, AWS and the rest divide the market, see our 2026 vendor landscape; for the product itself, the agent platform overview.

COEXISTENCE

Running Both Side by Side

Most Azure-first enterprises keep Foundry for Azure-native apps and add VDF AI where residency or portability rules apply.

1
Sort workloads by permitted location

Classify each agent use case by data class and by where it is allowed to execute. Workloads that may run in an Azure region can stay on Foundry; the rest need a landing zone you control.

2
Stand up VDF AI where Foundry cannot go

Install the self-hosted package in the data center, a sovereign tenancy or a disconnected enclave. Teams that want Azure infrastructure with a portable agent layer can run it on AKS instead.

3
Reuse what Microsoft already gives you

Connect SharePoint, Outlook and Azure DevOps through the Microsoft connector, which authorises with a work Microsoft account, and register your Azure OpenAI deployments as a model source beside local models.

4
Compare the evidence, not the demos

Run one regulated workflow on each platform and put both audit records in front of the reviewer who signed off your last compliance cycle. Their verdict is usually quicker than a feature matrix.

Map Your Azure Estate
FULL COMPARISON

Capability Table

Microsoft details taken from Microsoft Learn and azure.microsoft.com on 27 September 2026.

CapabilityVDF AIMicrosoft Foundry
Product typeGoverned agent platform, cloud or self-hostedAzure platform service for agents, models and tools
NamingVDF AI Agents, Networks, Router, ChatMicrosoft Foundry, formerly Azure AI Studio and Azure AI Foundry
Agent buildingAgents with Agent Skills; Networks for staged multi-agent workPrompt agents, voice agents (preview), hosted agents, workflows
Bring your own codeREST APIs across the platform for embedding and automationHosted agents run Agent Framework, LangGraph, OpenAI or Anthropic SDKs
Model choiceOpenAI, Anthropic, Azure OpenAI, Mistral, Ollama, OpenAI-compatible endpoints10,000+ catalog models; agent support varies by region
Models on your hardwarevLLM, Ollama and llama.cpp endpoints inside your networkFoundry Local on Azure Local (preview, by request)
Identity and accessSSO on Enterprise Cloud; Entra ID SSO on-premise; per-role tool grants in the MCP gatewayDedicated Entra agent identity, Azure RBAC, OBO
Tools and protocolsMCP gateway with per-role grants and per-call auditToolboxes via managed MCP endpoint; OpenAPI; A2A v1.0
GuardrailsAllowed models and services, budget caps, Human Approval nodesContent filters, prompt-injection mitigation, Azure Policy
Audit and evidenceVault record per run; EU AI Act evidence packsTracing, evaluations, Application Insights
DeploymentVDF Cloud, self-hosted, private cloud, air-gappedAzure regions; Azure Local options in preview
PricingPer user (Cloud); annual capacity pool (on-prem)Tokens, tool meters, hosted-agent vCPU and memory
Microsoft 365 reachConnector for OneDrive, Outlook, SharePoint, Planner, Azure DevOps, Dynamics 365Publish into Teams and Microsoft 365 Copilot

Preview features can change or disappear before general availability. Confirm region and feature status for your own subscription before you design around them.

Microsoft sources checked for this page

FAQ

Frequently Asked Questions

Questions Azure architects raise when they look past Foundry.

Yes. Microsoft Learn lists Azure AI Studio and Azure AI Foundry as the previous brand names and Microsoft Foundry as the current one (checked September 2026). Azure AI Services are now called Foundry Tools, and the older hub-based projects live on in the Foundry (classic) portal. The agent runtime inside it is Foundry Agent Service. Many buyers still search for the old name, which is why this page uses both.

Not as the same managed service. Prompt agents and hosted agents run on Microsoft-managed compute in Azure regions, with virtual-network isolation and bring-your-own storage as the isolation tools. For local sites Microsoft documents two separate Azure Arc extensions: Foundry Local on Azure Local, which serves models on Arc-enabled Kubernetes (preview, access by request), and Agentic Retrieval in Foundry Local, a local RAG agent layer with an MCP server (preview) that can run disconnected under the Azure Local disconnected operations preview. VDF AI ships its full platform as self-hosted packages, including an air-gapped install from your internal registry.

Microsoft says the Foundry platform is free to explore and that each feature you consume is billed at its own rate. Models bill per token. Prompt agents carry no extra agent charge beyond model tokens and tools. Hosted agents add compute billed per vCPU-hour and per GiB-hour of memory for each active session, and built-in tools such as file search storage, code interpreter sessions and web search transactions have their own meters. VDF AI charges per user on its Cloud plans and sells on-premises deployments as an annual capacity pool with unlimited users, where orchestration, routing, governance and audit logging are never metered.

Microsoft advertises more than 10,000 models in the Foundry catalog, from Microsoft, OpenAI, Anthropic, Meta and others, although Agent Service model and tool support varies by region. VDF AI connects agents to OpenAI, Anthropic, Azure OpenAI, Mistral, Ollama or any OpenAI-compatible endpoint, and to open-weight models served inside your perimeter through vLLM, Ollama or llama.cpp. Its router then picks a model per step from the list your administrators allow. Foundry wins on catalogue breadth; VDF AI wins when the model has to run on hardware you own.

Foundry gives each hosted agent its own Microsoft Entra identity at deploy time, supports On-Behalf-Of passthrough for user-delegated calls, and enforces access through Azure RBAC. That is a real strength inside an Entra tenant, and VDF AI has no equivalent per-agent directory identity. VDF AI governs access at the tool: administrators grant tools to roles in its MCP gateway registry, every tool call lands in the audit trail, and Human Approval gates in VDF AI Networks hold consequential steps for the accountable owner. Single sign-on is included on the Enterprise Cloud plan, and on-premise deployments sign in with Microsoft Entra ID natively, mapping Entra security groups to VDF AI roles. Those grants and records live in VDF AI, so they do not depend on an Azure subscription.

Microsoft positions Copilot Studio as its low-code SaaS agent builder and Foundry as the platform-as-a-service option for pro-code and declarative agents. Copilot Studio can call a Foundry agent through a connection that is still in preview. Microsoft Agent Framework is the open-source SDK that succeeds Semantic Kernel and AutoGen; you can run it anywhere, and Foundry hosted agents are one managed place to run it. If your team is weighing the low-code builder instead, our separate Copilot Studio comparison covers that decision.

Yes. Self-hosted VDF AI packages are containers that run on Docker Compose or on any conformant Kubernetes cluster, so AKS or plain virtual machines in your subscription both work, and the documentation names Azure Database for PostgreSQL as a supported managed database. You keep Azure networking, monitoring and procurement while the agent layer itself stays portable. The one practical difference is billing: VDF AI is bought directly or through a partner rather than through your Azure invoice.

Agent code written with open frameworks such as Agent Framework or LangGraph can move, and Foundry exposes OpenAI-compatible Responses endpoints. The surrounding services do not travel: Entra agent identities, Toolboxes, Foundry conversation storage, Application Insights traces and Azure Policy assignments are Azure resources. VDF AI keeps agents, networks, policies and audit records inside one product that runs the same way in its cloud, in your Azure tenant or on bare metal, so a hosting change does not force an agent rebuild.
Explore further

Related resources

Azure teams usually want to know how the Microsoft 365 estate, regulated workloads and model choice carry over, so these links start with the Microsoft connectors and end with deployment detail.

Need Agents Beyond Azure Regions?

Bring one workflow that cannot run in an Azure region and we will show it running in VDF AI on infrastructure you choose, still reading from Microsoft 365 through the Microsoft connector.

See Pricing Mechanics Book an Architecture Review