Documentationbreadcrumb arrow Grafana Cloudbreadcrumb arrow AI and machine learningbreadcrumb arrow Grafana Assistantbreadcrumb arrow Manage MCP serversbreadcrumb arrow Understand MCP servers
Grafana Cloud

Understand MCP servers

Learn how Model Context Protocol (MCP) servers integrate with Assistant to extend capabilities beyond observability data. Use this guide to understand authentication methods, server scope, tool filtering strategies, and lifecycle management.

What you’ll achieve

Use this guide to understand MCP server concepts and learn how to manage them effectively.

  • Connect external systems to Assistant for incident investigation and task automation.
  • Choose the right authentication method and scope for personal or team-wide access.
  • Filter tools to improve response speed and accuracy.
  • Manage server lifecycle including enabling, disabling, and removing configurations.

Before you begin

Ensure you have the necessary permissions and feature access before working with MCP servers.

Understand common use cases

MCP servers enable workflows that connect observability data with external systems for investigation, automation, and collaboration.

Common scenarios include:

  • Incident investigation: Search GitHub issues related to error messages, find recent code changes, and cross-reference incidents across systems.
  • Task automation: Create GitHub issues from investigation findings, update Linear tickets with root cause analysis, and link incidents to pull requests.
  • Knowledge retrieval: Search documentation repositories, find related incidents from your issue tracker, and retrieve context from external knowledge bases.
  • Collaborative troubleshooting: Share findings across team tools, update runbooks, and track resolution steps in your project management system.

MCP servers available to Assistant

You connect remote MCP servers to Assistant so it can call their tools during conversations.

Assistant supports these common integrations:

  • GitHub: Find related issues, search for code and errors, and propose small fix pull requests.
  • Linear: Search and manage issues and projects with safe, manual tool confirmations.
  • Custom integrations: Connect any compatible MCP server for specialized workflows.

MCP servers extend Assistant’s capabilities beyond metrics, logs, traces, and profiles by providing access to external systems and data sources. This lets you troubleshoot issues, investigate incidents, and complete tasks without leaving your conversation with Assistant.

Note

Assistant supports remote MCP servers. It doesn’t host or run local MCP servers.

Understand MCP server workflow

Assistant communicates with MCP servers through a standardized protocol to discover and invoke tools.

When you add an MCP server:

  1. Discovery: Assistant connects to the server and discovers available tools with their input schemas and descriptions.
  2. Filtering: You select which tools to expose to Assistant during conversations.
  3. Invocation: During a conversation, Assistant identifies when an MCP tool can help and proposes to use it.
  4. Confirmation: You review the tool call details and approve or reject the invocation.
  5. Execution: After approval, Assistant calls the tool through the MCP server and uses the results to continue the conversation.

This workflow ensures you maintain control over external actions while enabling Assistant to access data and perform operations across multiple systems.

Understand scope and limitations

You should be aware of server visibility, deployment requirements, and external provider constraints before configuring MCP servers.

Consider these limitations:

  • Personal vs tenant-level: Personal servers are visible to you; tenant-level servers are visible to all users in your organization with sufficient permissions. Choose personal scope for individual workflows and tenant scope for shared team resources.
  • Remote servers only: Assistant connects to remote endpoints; it doesn’t run local tool servers. The MCP server must be accessible from your Grafana instance over the network.
  • Rate limiting: The external provider enforces rate limits. Check the provider’s documentation for rate limit details and plan your tool usage accordingly.
  • Network access: Grafana Cloud must be able to reach the MCP server endpoint. Ensure firewalls and gateways allow outbound connections to the MCP server’s domain.
  • Tool availability: Not all MCP servers support the same tools. Review the discovered tools after connecting to understand what operations are available.

Choose the right authentication

Select an authentication method based on your scope and security requirements.

Authentication options include:

  • OAuth: Use the provider’s OAuth consent flow (supported for GitHub personal scope). This method redirects you to the provider’s authorization page where you grant permissions. OAuth tokens are scoped to your user account.
  • Custom auth header: Send a static header for authentication (required for tenant-level servers). Provide a service account token or API key that grants access for all users in your organization.
  • Unauthenticated: Use for public, read-only endpoints that don’t require authentication.

Warning

Tenant-level MCP servers must use a custom auth header, not OAuth. Provide a service token through a header to make the integration available to the entire tenant.

For personal MCP servers, use OAuth when available for better security and token management. For tenant-level servers, create a dedicated service account with appropriate permissions and use its token in a custom auth header.

Improve accuracy with tool filtering

Filter MCP server tools to improve speed and accuracy. Fewer tools reduce the model’s choice set and improve relevance. Start with the minimum set of tools you need and add more only when necessary.

Too many tools cause:

  • Slower responses: Assistant takes longer to evaluate which tool to use.
  • Lower accuracy: More tools increase the chance of Assistant selecting the wrong tool.

Evaluate tools based on frequency of use, relevance to your team’s workflows, safety (read-only versus write operations), and performance. Monitor Assistant’s behavior over a few conversations and remove tools it never uses. If Assistant indicates it lacks a capability, add the relevant tool back.

Confirm tool calls manually

Review and approve every tool invocation before Assistant executes it. Manual confirmation prevents unwanted actions and keeps you in control.

When Assistant proposes to use an MCP tool, you see:

  • Tool name: The specific tool Assistant wants to call.
  • Input parameters: The data Assistant sends to the tool.
  • Purpose: Why Assistant thinks this tool helps.

You can approve the tool call to proceed or reject it to continue without that action. This confirmation step is important for tools that create, update, or delete data in external systems.

Manage server lifecycle

Control server availability and configurations throughout their lifecycle. You can enable, disable, edit, or remove servers to control when and how Assistant uses them.

Enable or disable a server:

Temporarily stop or start using a server without deleting its configuration. When you disable a server, Assistant stops proposing tools from that server in new conversations, active conversations can no longer use the server’s tools, and the server configuration and tool filters remain saved for when you re-enable it.

When you enable a server, Assistant includes the server’s tools in new conversations and reconnects to the server to verify availability.

Use the disable feature to temporarily reduce active tools during high-load periods, test investigations with and without specific integrations, or control costs by limiting external API calls.

Edit a server:

Modify server configurations to update names or tool filters. You can change the server name or modify tool filters at any time, but you can’t change the URL or authentication method without removing and recreating the server. Changes apply immediately to new conversations.

Remove a server:

Permanently delete MCP server configurations when you no longer need them. When you remove a server, the configuration is permanently deleted, tool filters are lost, and Assistant stops proposing tools from that server. Before removing a server, verify no users are actively using it and consider disabling instead if you might need it temporarily.

Next steps