MCP connectors and tools in Azure SRE Agent

The problem: your agent can't see half your stack

Your agent has deep built-in access to Azure services - Application Insights, Log Analytics, Azure Monitor, Resource Graph. But your operational toolchain doesn't stop at Azure. Metrics live in Datadog. Security logs live in Splunk. Source code lives in GitHub.

During an incident, that gap forces you to manually bridge systems - querying Datadog in one tab, correlating with Azure Monitor in another, copy-pasting findings between tools. Your agent is powerful, but it only sees the Azure side.

How MCP connectors work

The Model Context Protocol (MCP) is an open standard for connecting AI agents to external tools. An MCP server wraps any service - a database, an API, a monitoring platform - and exposes its capabilities as tools your agent can discover and invoke.

When you connect an MCP server to your agent:

1. Auto-discovery - Your agent calls the server's tool listing and immediately knows what tools are available.
2. Namespaced registration - Each tool is registered with a prefixed name (for example, my-datadog_list_metrics) so tools from different servers never conflict.
3. Health monitoring - Your agent pings each server every 60 seconds, auto-recovers from transient failures, and shows real-time connection status.
4. Dynamic updates - When a server adds new tools, your agent detects them automatically within 5 minutes.

Two transport types

MCP supports two ways to connect your agent to external tools:

Streamable-HTTP (remote services)

Connect to any MCP server accessible through a URL. Provide the endpoint and authentication credentials - either a Bearer token or custom headers.

Stdio (local processes)

Run an MCP server as a process inside your agent's cloud environment. Provide the command, arguments, and optional environment variables. Your agent manages the process lifecycle - starting it on connection, monitoring health, and restarting on failure.

Preconfigured partner connectors

Your agent includes preconfigured connectors for popular platforms. These connectors have prefilled URLs, default authentication settings, and branded icons. Select the card, enter your credentials, and connect.

For partner connectors, the authentication method is preconfigured and locked. You just enter the credential values.

How tools reach your agents

MCP tools can reach your agent in two ways:

• Agent tools: Select tools during connector setup or editing. These tools are available directly in the main conversation. No custom agent needed.

• Custom agent tools: Assign tools to specific custom agents for focused specialists with domain expertise.

Agent tools

When you create or edit an MCP connector, a tool selection step lets you choose which tools are visible to your agent. The selected tools are added to your agent's tool list and stay in sync. When you add, remove, or update connectors, your agent's tools refresh automatically.

During connector creation: After your connector connects successfully, a Select tools step appears. All discovered tools are preselected up to the remaining capacity. Select Done to save or Skip to add tools later.

For existing connectors: Edit any MCP connector to find the MCP Tools section at the bottom of the dialog. Check or uncheck tools to control which ones your agent can call.

Custom agent tools

For focused specialists, assign MCP tools to specific custom agents through the portal or YAML.

Portal: In Builder > Agent Canvas, edit an agent → Advanced settings > Tools → Choose tools.

YAML wildcards:

mcp_tools:
  - datadog-mcp/*        # All tools from this connection
  - github_search_code   # One specific tool

The {connection-id}/* pattern adds every tool from that server, including tools added later. The forward slash is required.

For a full walkthrough, see Set Up MCP Tools.

Tool selection and capacity

Each agent - whether your agent or a custom agent - can use up to 80 tools (native and MCP combined). The tool selection UI helps you manage this budget across connectors.

Capacity indicator

The tool picker shows a progress bar with your current tool count:

A hint below the bar reads: "Maximum of 80 tools allowed to avoid degrading agent performance."

Behavior at capacity

When your selected tools reach the 80-tool limit:

• Unchecked tools become disabled - you can't add more until you remove some
• Already-checked tools stay enabled - you can always deselect them to free capacity
• Select all caps at remaining capacity - if you have room for 10 more tools and select Select all, only 10 are selected

The tool count spans all connectors. If Connector A has 60 tools visible to your agent, Connector B can add up to 20 more.

Connection health monitoring

Your agent continuously monitors every MCP connection:

Your agent pings each server every 60 seconds. Transient failures recover automatically on the next successful heartbeat. Before invoking any MCP tool, your agent validates the connection and attempts reconnection if needed.

Auto-reconnection

Your agent automatically recovers from MCP connection failures by using two mechanisms:

• Before every tool call, your agent checks the connection status. If the connection is disconnected, it reconnects transparently before executing the tool - no waiting for the next heartbeat cycle.

• Every 60 seconds, your agent pings each Streamable-HTTP MCP server. If a disconnected server responds to the ping, the connection recovers. This mechanism catches failures that happen between tool calls. Stdio connections recover on the next tool call rather than through heartbeat pings.

When a connection reconnects:

• Your agent establishes a new session with the MCP server.
• Your agent re-discovers all available tools from the server.
• Your agent refreshes tool references so every subsequent call uses the live connection.
• Existing tool assignments (agent tools and custom agent tools) remain intact.

Authentication

Each MCP server has its own authentication requirements. The portal supports the following methods:

For partner connectors, the auth method is preconfigured - you just enter the credentials.

What makes this different

• Unlike custom integrations, MCP connectors use an open standard. You don't write adapter code - any MCP-compatible server works out of the box. The protocol handles discovery, invocation, and error handling.

• Unlike static tool configurations, your agent discovers tools dynamically. When an MCP server adds a new tool, your agent picks it up automatically. Wildcard patterns (datadog-mcp/*) ensure new tools are available without reconfiguration.

• Unlike scripts and runbooks, MCP connections are self-healing. Your agent monitors every connection with a 60-second heartbeat, auto-recovers from transient failures, and defers custom agent loading until connections establish.

Before and after

Related capabilities