Product Launch

Microsoft Foundry Reveals MCP Firewall to Screen Agent Calls

Microsoft Foundry's Dev Pack and MCP Firewall preview debut at Ignite, giving enterprise teams governance controls over every AI agent tool call.

Share:XLinkedIn

Key Takeaways

  • Foundry Dev Pack reaches GA, bundling agent tools for VS Code, the terminal, and compatible coding agents with Foundry Toolboxes for sharing tools through a managed MCP endpoint across teams
  • Entra MCP Firewall enters preview with network-level controls over which MCP servers AI agents can connect to, backed by Entra identity and audit trails for regulated enterprises
  • Copilot Fusion Max scored 71.2% on Terminal-Bench 4.1, 5.4 points ahead of Anthropic's Opus 5.5 and Fable 5.2, at 88% lower cost than Fable 5.2 per GitHub's official changelog
  • Agent 365 MCP governance enters Frontier Public preview with spend visibility, guardrail enforcement, and ROI tracking through LiteLLM and Azure AI Gateway integration
  • More than 100,000 apps have been built with Rayfin since preview, giving Microsoft a concrete adoption data point as the Foundry ecosystem competes against Salesforce Agentforce

Microsoft's most consequential AI announcement at Ignite 2026 didn't come from its model team. It came from its identity team, in the form of a network firewall built specifically for the Model Context Protocol. The Entra MCP Firewall, entering preview on November 19, is the first enterprise-grade product designed to control which MCP servers AI agents are permitted to connect to. Alongside it, the Foundry Dev Pack reached general availability, bundling Microsoft's agent development tools into a single package for Visual Studio Code, the terminal, and compatible coding agents.

What the Foundry Dev Pack Actually Bundles for VS Code Developers

The Foundry Dev Pack is now generally available. Its design is straightforward: Microsoft is packaging its agent development scaffolding into one installable bundle rather than requiring developers to assemble it themselves. The pack covers Visual Studio Code, the terminal, and coding agents that implement the Agent Control Protocol, according to the Microsoft Ignite 2026 Book of News.

The centerpiece is Foundry Toolboxes, which let engineering teams define a shared set of tools and distribute them through a managed MCP endpoint. Where developers previously had to configure individual MCP server connections across projects and teammates, Toolboxes centralizes that into a single org-level endpoint controlled by administrators. The goal is to make tool access as manageable as a software dependency, not as ad hoc as a configuration file scattered across a team's machines.

Alongside the Dev Pack, the Microsoft Agent Framework, which reached 1.0 in April 2026, provides the runtime underneath. It supports connections to more than 11,000 models through Foundry, including third-party providers from Anthropic, Google, Amazon Bedrock, and Ollama. For developers who want a single SDK covering both single-agent and multi-agent patterns, the framework now offers stable APIs and a long-term support commitment across .NET and Python. The Copilot SDK 2.0 also reached general availability at Ignite, shipping a Rust runtime across six languages, reducing the memory footprint of agent processes running in resource-constrained edge environments.

Stay Ahead

Get daily AI signals before the market moves.

Join founders, investors, and operators reading TechFastForward.

One feature still in preview is hosted-agent endpoints, which would let Foundry builders deploy agents without managing their own infrastructure. That is the gap Microsoft is asking developers to tolerate while the governance layer catches up. The Visual Studio Magazine's coverage from Day 3 of Ignite specifically noted the Dev Pack GA as the most immediately actionable announcement from the conference for developers building agent-powered applications, distinguishing it from the governance tools that remain in preview.

Why MCP Security Became Microsoft's Highest Priority Bet

The Model Context Protocol, introduced by Anthropic in late 2024, is by design a permissive standard. Any agent that speaks MCP can connect to any MCP server that accepts its credentials. For a developer building a local coding tool, that's a feature. For an enterprise running hundreds of agents across its entire software stack, it's a liability that grows with every new integration.

The MCP ecosystem grew faster than anyone anticipated. By mid-2026, more than 500 third-party MCP servers were listed in community registries, covering JIRA, Slack, Stripe, Salesforce, database connectors, and payment APIs. The speed of that growth outpaced any attempt at standardized governance. Security researchers began flagging a concrete risk: an agent with broad MCP access and an imprecise system prompt could exfiltrate data by calling an unintended server in response to a seemingly benign request. Microsoft's own security research team published findings earlier in November documenting that threat actors had already used agentic tooling to identify exposed credentials in public repositories, a preview of what happens when agents have unchecked tool access at scale.

The Entra MCP Server, now generally available, addresses the authentication half of the problem. It provides enterprise identity for MCP connections using Microsoft Entra, the same identity service that governs human logins across Microsoft 365, Azure, and thousands of enterprise applications. When an agent calls an MCP server via Entra MCP, the connection is authenticated, the identity is logged, and the access is auditable. This matters for compliance teams: the first question any regulated enterprise asks about an AI agent is whether it has an audit trail. The Entra MCP Server makes the answer yes, according to the Microsoft Copilot blog.

The Entra MCP Firewall goes further. In preview as of November 19, it adds network-level policies: administrators can specify which MCP servers enterprise agents are permitted to connect to, block connections to servers outside approved categories, and enforce these controls at the organization level rather than leaving them to each team's judgment. Agent 365 MCP governance, also entering preview, integrates with LiteLLM and Azure AI Gateway to give platform teams visibility into which agents are calling which tools, how much it costs, and whether calls fall within compliance guardrails. The implications for independent MCP server providers are significant: more than 500 MCP server providers in community registries may need to adopt Entra to maintain enterprise adoption in regulated industries. If Entra MCP becomes the enterprise standard for authentication, server providers who want adoption in regulated industries will need to integrate Entra, not just implement the base MCP spec. This is a playbook Microsoft has run before with OAuth and Active Directory, extending a standard it did not create into infrastructure it controls.

The 71.2% Score That Puts Copilot Fusion Into the Cost War

The governance story at Ignite runs alongside a more direct competitive argument: Microsoft is claiming its hosted coding model costs less than the alternatives while matching them on third-party benchmarks.

Copilot Fusion Max scored 71.2% on Terminal-Bench 4.1, a coding benchmark maintained on public leaderboards at Stanford, Harbor, and the Laude Institute. That puts it 5.4 percentage points ahead of Opus 5.5 and Fable 5.2, the two Anthropic frontier models it benchmarks against. On cost, the GitHub Copilot changelog reports Fusion Max is 88% cheaper than Fable 5.2 and 57% cheaper than Opus 5.5. The broader cost claim, from a GigaOm study cited by Microsoft, is that Copilot is 84% cheaper than Claude and 63% cheaper than ChatGPT for a typical enterprise user.

The Terminal-Bench 4.1 leaderboard tests coding agents on real-world terminal operations: file manipulation, build systems, package management, and debugging. It is not a purely synthetic benchmark, which makes it somewhat more predictive of production use than purely mathematical evaluations. Microsoft's decision to report Fusion's score on a neutral third-party leaderboard rather than an internal evaluation is deliberate; it expects the number to withstand comparison by practitioners who already track these rankings.

Fusion is not a single model but a routing system that selects from Claude Sonnet 5.5, GPT-6.1 Terra, and MAI-Code-2 based on task type. The benchmark measures full-system output, not any single underlying model in isolation, which makes direct comparison to standalone Anthropic or OpenAI model scores structurally different. What it measures is the system an enterprise actually deploys, which is arguably the right unit of comparison. Fusion is now available in the model picker on GitHub.com, Visual Studio, and GitHub Mobile beta apps for iOS and Android. Production mobile support is "coming shortly," suggesting the rollout is still completing.

Autopilot, Agent 365, and What 100,000 Rayfin Apps Signal

Copilot Autopilot is rolling out today to Frontier program customers. It's designed for recurring, long-horizon work: maintaining state across sessions, picking up unfinished tasks, and handling the kind of repetitive structured work that fills engineering and operations team backlogs. The Frontier program, Microsoft's early-access tier, is where the company routes customers who can tolerate rough edges in exchange for access to features before they reach general availability.

Agent 365 Business Analytics is now in Frontier Public. It aggregates spend data, usage by agent type, adoption curves, and ROI signals across an organization's deployed agents. The Cost Management and FinOps for AI components give finance and platform teams something they haven't had before: a central place to see what their AI agents cost in aggregate, not just per-API-call. This matters because agent billing is spread across compute, model calls, tool calls, and storage, and organizations are frequently surprised by aggregate costs when they're watching only line-item expenses.

One of the more striking data points from the Book of News is that more than 100,000 apps have been built with Rayfin since its preview launch. Rayfin, Microsoft's low-code agent builder, competes with Salesforce Agentforce and similar products in the no-code and low-code enterprise segment. That adoption number, while company-reported and unaudited, suggests real developer uptake rather than conference-announcement interest alone. The metric gives Microsoft a concrete counter to the narrative that enterprise agent adoption has stalled behind governance concerns.

Azure Catalyst, now generally available, rounds out the developer stack with agent governance rulebooks inside GitHub Copilot's application-modernization workflows. It also adds bring-your-own-model support in the CLI and IDEs, letting enterprises route specific agent tasks to models outside the Microsoft ecosystem while keeping the governance and observability layer intact. This is an important concession to multi-cloud reality: rather than forcing enterprises to standardize on Microsoft models, Catalyst is designed to work as a governance wrapper around whatever models an enterprise already runs.

Microsoft also shipped Microsoft-Decision-1.1 at Ignite, a model built on Gemma 4 26B A4B that scored 68 on the Artificial Analysis Decision Index v2.1. Decision models are optimized for structured choices rather than generation tasks, and Microsoft has been building them specifically for agent orchestration scenarios where agents need to decide between competing actions rather than produce natural language output. The model's GA alongside everything else at Ignite signals that Microsoft expects the full agentic stack to require specialized components, not just a single generalist model handling every task.

The Bear Case: Why Governance Might Arrive Too Late

The gap between Ignite announcements and production-ready software is not small. Of the developer tools announced across the three days ending November 19, the Foundry Dev Pack and Copilot SDK 2.0 are the only components that are generally available. The Entra MCP Firewall, Agent 365 MCP governance, hosted-agent endpoints, and Windows 365 for Agents are all in preview, with varying timelines to GA.

Critics argue that shipping preview governance tools alongside GA developer tools creates a window where development runs ahead of compliance, potentially giving organizations a liability gap they didn't anticipate. The pattern is familiar from Microsoft's cloud migration era, when Azure services frequently launched in preview with the implicit expectation that enterprises would treat them as production-ready anyway. When the governance pieces eventually reach GA, they arrive into an environment where agents have been running ungoverned for months.

The competitive threat is real as well. Anthropic released enterprise governance primitives for Claude earlier in 2026, and OpenAI's enterprise APIs have included role-based access control, logging, and content policies for some time. Neither requires Microsoft's identity infrastructure to function. If Anthropic or OpenAI ship a comparable MCP governance layer before Microsoft's Entra MCP Firewall reaches GA, the enterprise argument for staying within Microsoft's ecosystem weakens. The bear case is simply this: Microsoft's governance strategy requires enterprises to wait for preview features while their agents are already in production on stacks that have no governance layer at all.

There is also the question of fragmentation. The MCP spec is model-agnostic and controlled by Anthropic through an open process, with contributions from Google, OpenAI, and other vendors. An Entra-dependent MCP Firewall is specifically Microsoft's extension of that spec, not a contribution to it. Developers building multi-cloud agent deployments, including those running agents on AWS Bedrock, Google Vertex, or Anthropic's direct API, would need to implement Microsoft's specific authentication layer to benefit from Foundry governance. That's a concrete adoption hurdle for any organization not already running primarily on Azure. Coverage from Witivio and Exelegent both note that the most immediately actionable item from Ignite for the broader developer community is the Fusion cost reduction, which requires no configuration change, while the governance story is a longer adoption cycle.

Three Signals to Watch Before End of November

Three specific indicators will tell whether this Ignite release lands as Microsoft's enterprise control plane moment or as an impressive preview dump.

First: Copilot Code, Microsoft's on-device coding runtime, goes "generally available in the coming weeks," according to the Copilot blog published November 17. That window is now narrow, and the precise date will indicate whether Microsoft is on the aggressive timeline it implied or whether it's slipping into Q1. Copilot Code is the missing piece between the Foundry Dev Pack and Autopilot's recurring-workflow story: without GA code execution inside the agent loop, Autopilot's value is limited to orchestration, not full automation. Watch for a release note on GitHub's changelog by November 30.

Second: the Entra MCP Server's enterprise adoption rate in the first 30 days. The server is generally available, and it's the foundation that the MCP Firewall preview is built on. If adoption among major MCP server providers, specifically Stripe, Atlassian, and Salesforce, moves quickly, it validates Microsoft's thesis that enterprise authentication will be the chokepoint that consolidates the MCP ecosystem. If adoption is slow, it suggests that server providers are betting on a neutral cross-vendor auth standard rather than accepting Microsoft's implementation as the standard. Atlassian's public response will be the leading indicator to watch, since Atlassian serves both GitHub and non-GitHub enterprise customers.

Third: Anthropic's response to the 84% cost comparison claim. Claude models are simultaneously a component in Microsoft's Fusion routing system and the benchmark Fusion is claiming to beat. Anthropic's enterprise relationship with Microsoft is structurally unusual: Anthropic models are components in a product that competes against Anthropic's own direct channels. How Anthropic chooses to respond to the GigaOm cost comparison, whether by commissioning a counter-study, by cutting Claude API prices, or by staying silent, will shape the developer market dynamic heading into Q1 2027. The specific date to watch is Anthropic's next model pricing announcement, expected before year-end.

Microsoft's MCP Firewall is not a product most developers will use on day one, but it is the product that will determine whether enterprises trust AI agents with their most sensitive tool calls in 2027.


Key Takeaways

  • Foundry Dev Pack reaches GA, bundling agent tools for VS Code, the terminal, and compatible coding agents with Foundry Toolboxes for sharing tools through a managed MCP endpoint across teams
  • Entra MCP Firewall enters preview, giving enterprise administrators network-level controls over which MCP servers AI agents can connect to, backed by Entra identity and audit trails
  • Copilot Fusion Max scored 71.2% on Terminal-Bench 4.1, 5.4 points ahead of Anthropic's Opus 5.5 and Fable 5.2, at 88% lower cost than Fable 5.2 per GitHub's changelog
  • Agent 365 MCP governance enters Frontier Public with spend visibility, guardrail enforcement, and ROI tracking through LiteLLM and Azure AI Gateway integration for enterprise platform teams
  • More than 100,000 apps have been built with Rayfin since preview, giving Microsoft a concrete adoption data point as the Foundry ecosystem competes against Salesforce Agentforce and similar platforms

Questions Worth Asking

  1. If Entra MCP becomes the enterprise standard for agent authentication, do open-source MCP server maintainers gain leverage over Microsoft's integration requirements or lose the ability to stay cross-platform?
  2. At 88% cheaper than Fable 5.2, is Copilot Fusion the first agentic coding system where the cost savings alone justify switching regardless of benchmark differences?
  3. When governance tools enter preview 12 to 18 months after the agents they govern are already in production, can enterprises apply them retroactively without disrupting live workflows?

Sources

  1. Microsoft Ignite 2026 Book of News news.microsoft.com
  2. Microsoft Agent Framework devblogs.microsoft.com
  3. Microsoft Copilot blog microsoft.com
  4. GitHub Copilot changelog github.blog
  5. Witivio witivio.com
  6. Exelegent exelegent.com

Read Next

SpaceX Sells $32B in Bonds to Finance Nvidia Rubin Chips

Anthropic Raises $119.5B in Record IPO, Stock Soars 17%

Democrats Demand DOJ Records on OpenAI Copyright Filing

DJI Air 3S Outranges Neo by 25 km and Costs 5.5x More

Newsletter

Enjoyed this analysis? Join the list for the email edition.

Daily AI signals. No noise. Built for founders, investors, and operators.

Share:XLinkedIn
</> Embed this article

Copy the iframe code below to embed on your site:

<iframe src="https://techfastforward.com/embed/microsoft-foundry-reveals-mcp-firewall-screen-agent-calls" width="480" height="260" frameborder="0" style="border-radius:16px;max-width:100%;" loading="lazy"></iframe>