A glowing cloud-based automation hub connects data tools to research and deployment workflows staffed by robot assistants.
Microsoft used Ignite to widen Foundry Toolboxes from a tidy idea into a full platform pitch. The idea is simple: stop wiring the same integrations into every AI agent, and build them once. Define a curated set of tools, skills and agents, version it, and publish it behind one MCP endpoint. Foundry agents and other MCP-compatible runtimes then consume it.

The announcement is long, so this piece separates what is generally available, what is in preview, and what Microsoft only says is "coming soon." The shipping status matters more than the pitch when you plan a rollout.

A glowing cloud-based automation hub connects data tools to research and deployment workflows staffed by robot assistants. The core idea: one endpoint, many agents​

Microsoft frames the problem as fleet management, not connecting one tool to one agent. Every new agent otherwise repeats the same integrations, copied instructions and dependency troubleshooting. A toolbox centralizes downstream authentication, so agents connect to the toolbox and don't hold credentials for each tool.

Microsoft Learn's documentation adds detail on the mechanics:

  • There are two endpoint patterns. A developer endpoint targets a specific version (/toolboxes/{name}/versions/{version}/mcp) for testing before promotion. A consumer endpoint has no version in the path and is the one agents connect to.
  • The first version becomes the default, and creating later versions does not switch the default automatically. So you can stage a change, test it, and promote it when ready.
  • Microsoft's MCP documentation says versioning lets you iterate on a new version without affecting production agents.

The practical result is a split between testing and production. Validate v2 on the version-specific endpoint, then promote it. Agents on the consumer endpoint pick it up without code changes. That is a real operational benefit, and it is also a single point of failure. A bad promotion reaches every consuming agent at once.

Tool search keeps big catalogs usable​

A shared catalog that grows into dozens or hundreds of tools would bloat every model request. Tool search is Microsoft's answer. Per Microsoft's docs, it hides tools by default and exposes two meta-tools: tool_search to find relevant tools, and call_tool to invoke one by name. Administrators can pin critical tools, add organization-specific discovery terms, and auto-pin frequently used tools.

You can check that it is working. Microsoft's tool-search documentation says to use the version-specific endpoint and confirm that tool_search appears in tools/list and that no other toolbox tools show in the initial listing. If you see your whole catalog in the first response, tool search is not active on that version.

What was announced, and its status​

CapabilityStatus in the announcement
Skills CRUD APIGA "coming soon"
Skills via Shell tool in prompt agentsPublic preview
Skills exposed via official MCP Skills extensionDescribed as upcoming
Web Search powered by Web IQAvailable in prompt agents; Toolbox "coming soon"
Fabric IQ tool (60+ Fabric item types)Preview
Work IQ API (Web IQ and Foundry IQ access)Preview; Fabric IQ to follow
A2A tool and A2A endpoints (A2A protocol 1.0)Tool GA; endpoints GA on prompt agents; hosted-agent endpoints public preview, coming soon
Snowflake, MongoDB, Databricks in tools catalogNew catalog partners
One-click Toolbox install in GitHub Copilot in VS CodeAnnounced as available now
Toolbox monitor dashboardAnnounced
Tool governance with Azure PolicyPublic preview
Browser automation toolGA; VNet injection and several extras in preview
Computer Use toolGA; Windows 365 for Agents MCP is a limited preview

Skills, Shell and the Web IQ search upgrade​

Skills package task-specific guidance, such as analyzing business data, preparing customer briefings, or checking documents against company standards. Microsoft's docs describe them as versioned and immutable, and a toolbox can pin a specific version. They are still labeled preview, and the Skills CRUD API is not yet GA.

The Shell tool gives prompt agents a managed environment to load a skill and run the needed code. Microsoft's docs also note that skills attached to a toolbox must exist in the same Foundry project, and cross-project references aren't supported. A client that wants to consume skills over MCP needs support for the MCP Skills extension. Reading a resource alone doesn't activate a skill.

General Web Search is now powered by Web IQ. Microsoft says it returns focused, citation-ready context and needs no separate Web IQ connection setup. The Toolbox side is "coming soon," so don't plan a shared Web IQ search tool today. For domain-restricted searches, Grounding with Bing Custom Search still applies.

Fabric IQ, Work IQ and permission-aware context​

The Fabric IQ tool connects agents to Power BI semantic models, ontologies and Fabric data agents. In preview, it can discover more than 60 Fabric item types, which developers can browse and filter. Microsoft's example is an operations agent that explains a spike in checkout failures using metric definitions and permissions already set in Fabric. Existing Power BI semantic model and Fabric data agent integrations also get lower latency and clearer query-based tracing.

The Work IQ API, in preview, provides access to Web IQ and Foundry IQ, with Fabric IQ to follow. Microsoft describes it as spanning Microsoft 365 knowledge plus Dynamics 365 and Power Apps data, and as permission-aware. These are vendor claims. The announcement does not show how permission trimming behaves in edge cases, so test it with restricted users before trusting it.

Agent-to-agent and the partner catalog​

The A2A tool and endpoints use the stable 1.0 version of the A2A protocol. A Foundry agent can delegate to another Foundry agent or an external A2A-compatible one. Each team keeps ownership of its domain agent and release cycle, while others call it. Microsoft's incident-response example has an on-call agent collecting evidence and asking a business-impact agent for assessment.

Snowflake, MongoDB and Databricks join the tools catalog. The catalog lets teams review each partner's setup and authentication requirements. The partner quotes in the announcement come from Databricks, Snowflake and MongoDB executives. Treat them as vendor statements, not independent evaluations. Their shared message is that existing permissions and business definitions still govern data access.

GitHub Copilot in VS Code​

For developers, this is the most immediately visible piece:

  1. In Foundry, open the Toolboxes tab and choose Install in VS Code.
  2. Confirm the installation in VS Code. The toolbox's MCP endpoint is added without copying URLs or editing configuration files.
  3. If authentication is needed, sign in with the Microsoft Entra account you use for Foundry.

Copilot connects as that user rather than through a shared toolbox credential, and each underlying tool keeps its own authentication and access controls. With Tool Search enabled, Copilot finds and calls tools on demand, and developers can inspect those searches and calls in chat. This is a meaningful governance point. Per-user identity preserves audit trails, but you still have to configure access on each underlying tool.

Monitoring and policy​

Microsoft calls shared tools production dependencies, and it is right. The monitor dashboard shows call volume, success rates, error trends, active agents, tool selection, skill and A2A usage, and consumption. You can filter by toolbox version, time range and agent, then drill into the consuming agent's traces.

Tool governance with Azure Policy is in public preview. Administrators can specify allowed tool types, MCP endpoints and MCP connectors. Evaluation happens when agents or toolboxes are created or updated, and when tools are invoked through them. Microsoft's advice is to start with Audit, which flags noncompliant configurations without blocking saves. Move to Deny once you understand what would break.

Browser automation and Computer Use​

The Browser automation tool, powered by Azure Playwright Workspaces, is generally available. It gives each agent a managed, isolated browser. Session-recording observability is GA, and private website access through VNet injection is preview. Further preview items include a bring-your-own browser proxy, Azure Key Vault secret management so the model doesn't see secrets, and CAPTCHA and Web Bot Auth handling. Profile reuse is "coming soon."

Web Bot Auth, based on an IETF draft, lets agents cryptographically sign HTTP requests so sites can verify the agent. Customers can self-register. Alternatively, they can express interest in Microsoft platform registration via bot-management providers such as Akamai, Cloudflare and F5.

The Computer Use tool is GA with supported OpenAI computer-use models. Customers bring their own harness and target application. For Windows admins, the notable item is a streamlined way to attach the Windows 365 for Agents MCP so agents can run computer-use tasks on governed Cloud PCs. That one is a limited preview, with an interest form.

Hands-on sample​

Microsoft's sample builds a shared troubleshooting toolbox in Python. It requires a model deployment that supports the Shell tool.

  1. Create a versioned skill named ops-analysis. Its instructions tell the agent to gather metrics, incident details and deployment history, compare equivalent time windows, and separate facts from hypotheses. They also say to cite sources, ask for missing data instead of inventing it, and recommend changes without making them unapproved.
  2. Create a toolbox version, ops-analysis-starter. It bundles Microsoft Learn's MCP server, Web Search, a Shell tool mounting the skill, and Tool Search. The code uses AIProjectClient with DefaultAzureCredential and allow_preview=True, and prints a version-specific MCP endpoint.
  3. Create a prompt agent, ops-troubleshooter, with an MCP tool pointing at that endpoint. Ask it to find Microsoft Learn guidance for high latency in Azure App Service and summarize three checks with citations.

Two cautions apply:

  • The sample sets require_approval="never" on its MCP connections. That keeps the demo smooth. It is not a hardening recommendation. Microsoft's docs say that when a tool is set to "always," the agent runtime must ask the user and wait for confirmation. Enforcement is the runtime's job, not the MCP endpoint's.
  • The sample's printed endpoint is version-specific. Microsoft's docs recommend that production agents use the unversioned consumer endpoint so promotions flow through. Reserve the version-specific one for testing.

Gotchas from the documentation​

  • Microsoft's toolbox guide says a toolbox supports at most one unnamed instance of Web Search, Foundry IQ (Azure AI Search), Code Interpreter or File Search. Two unnamed instances of the same type return an invalid_payload error. Give each a unique name.
  • Tool availability depends on region and model. Check support before you build around a tool.
  • Microsoft warns that non-Foundry tools may incur costs, and data may leave Foundry's compliance boundary.
  • If tools/list returns zero tools for MCP or A2A tools, the docs point to invalid or missing credentials for the remote server. Verify the project connection and RBAC.
  • Add a description to every tool. It helps the model pick the right one.

Analysis: who should care, and what to watch​

The strongest case is for organizations already running several agents and tired of duplicated plumbing. Central credentials, versioned promotion, per-user identity in Copilot, and Azure Policy allowlists address real enterprise pain.

The caveats are worth stating plainly:

  • Preview sprawl. Skills, Shell, Fabric IQ, Work IQ, tool policy and several browser features are preview. Web Search in Toolbox and the Skills CRUD GA are still "coming soon." Preview terms rarely suit production SLAs.
  • Shared blast radius. One toolbox serving many agents makes a promotion or a failing tool a fleet-wide event. The monitor dashboard and version pinning exist for this reason. Use them.
  • Vendor-sourced claims. Performance and "grounding quality" statements about Web IQ come from Microsoft, with no benchmarks in the announcement. Likewise, the partner quotes are marketing.
  • Not a security substitute. Each underlying tool still needs correct authentication and least-privilege access.

A sensible plan follows Microsoft's own closing advice. Pick one repeated workflow and publish it as a versioned toolbox. Point a second agent at the same consumer endpoint, then watch the dashboard for failures. Turn on the Azure Policy approved-tools rule in Audit mode before you let the catalog grow. Pin versions where you need controlled rollouts.

Toolboxes won't make agents smarter, but they might make fleets of them less annoying to run.

 

References

  1. Build once, reuse across your agent fleet: Tools and Toolboxes in Foundry Microsoft Foundry Blog 2026-11-18T17:00:18+00:00
  2. azure-ai-docs/articles/foundry/agents/how-to/tools/tool-search.md at main · MicrosoftDocs/azure-ai-docs · GitHub github.com
  3. Microsoft Foundry Toolbox and Tool Search: One Endpoint for Agent Tool Governance - DEV Community dev.to