SRE Agent MCP Server

TL;DR

Manage your SRE Agent from any MCP-compatible client, including VS Code, Copilot CLI, Cursor, Claude Desktop, and custom agent frameworks
Start investigations, configure connectors, search memories, and manage scheduled tasks in natural language
Two RBAC roles required: Reader (ARM) + SRE Agent Administrator (data plane)
No agent-side configuration needed. The Azure MCP Server discovers agents dynamically.
Secrets are redacted before reaching your client. Destructive operations require explicit confirmation.

Your agent's expertise, wherever you work

Your SRE Agent builds deep context about your services over time: incident patterns, architecture details, operational expertise. With the SRE Agent MCP Server, that knowledge is available directly in your IDE, terminal, or AI assistant.

During coding, debugging, or incident response, you can ask your agent a question, start an investigation, or configure a connector without switching tools. Your development environment and your agent's operational intelligence connect through the same natural language interface you already use.

How to connect

The SRE Agent tools are part of the Azure MCP Server, which implements the Model Context Protocol (MCP). You install the Azure MCP Server in your MCP client, and the SRE Agent tools become available alongside other Azure tools. The server runs locally via npx and handles authentication, endpoint resolution, and API calls on your behalf.

Two API layers handle different operations:

Layer	What it handles	Authentication
Control plane (ARM)	Agent resources, connectors	Reader role via Azure Resource Manager
Data plane	Threads, memories, tasks, skills, incidents	SRE Agent Administrator role via agent endpoint (`*.azuresre.ai`)

The server resolves agent endpoints automatically through Azure Resource Graph. You provide a subscription and agent name, and the server finds the endpoint.

Tools appear with the sreagent_ prefix in your MCP client (for example, sreagent_agents_list, sreagent_threads_create).

Permissions

Two Azure RBAC roles on the Microsoft.App/agents resource:

Role	Scope	What it enables
Reader	Control plane (ARM)	List and get agents, connectors
SRE Agent Administrator	Data plane	Threads, memories, scheduled tasks, skills, hooks, prompts, incidents, workflows

Supported clients

Client	Installation method
VS Code with GitHub Copilot	Install the Azure MCP Server extension, sign in to Azure
GitHub Copilot CLI	`/mcp add` or manual config in `~/.copilot/mcp.json`
Cursor	Add to MCP configuration
Claude Code	Add to user or project MCP configuration
Claude Desktop	Install MCPB bundle or configure local server command
Other MCP clients	Configure with `npx`, `dotnet`, `uvx`, Docker, or other supported methods

What you can do

Key capability areas, each accessible through natural language prompts:

Area	Operations	Example prompt
Manage agents	List, get details, create/delete sub-agents	"List my SRE agents in subscription X"
Configure connectors	Create Kusto/MCP connectors, test, delete	"Create a Kusto connector named prod-logs on agent Y"
Run investigations	Create threads, send messages, autonomous investigation	"Investigate why production API has elevated latency"
Schedule work	Create, pause, resume, delete scheduled tasks	"Pause the nightly scheduled task on agent Y"
Manage incidents	List active incidents, set up PagerDuty/ServiceNow	"List active incidents on agent Y"
Knowledge and prompts	Search/upload memories, manage common prompts	"Search memories for 'deployment failures'"
Author workflows	Generate, validate, apply YAML workflows	"Generate a workflow for automated rollback"

Autonomous investigation

The investigate command runs a multi-turn investigation loop. Your agent reasons about the problem, requests data, forms hypotheses, and follows up automatically.

Default limits: 20 iterations, 10-minute timeout (both configurable via --max-iterations and --timeout-seconds)
Standard mode: Pauses at approval gates for human confirmation
Auto-approval mode (investigate_yolo): Continues through all gates autonomously

Use auto-approval mode with extreme caution

The investigate_yolo command auto-approves all approval gates, including actions that modify your infrastructure (pod deletion, Kubernetes YAML application, scaling, incident state changes). There is no read-only restriction. The agent may invoke any tool its managed identity permits. Do not use in production unless you accept fully autonomous infrastructure modification.