Skip to main content

Overview

MCP (Model Context Protocol) lets AI SRE agents connect to external tools and data sources. Each MCP server is a standardized endpoint that exposes a set of tools (functions). Examples include querying GitHub issues, sending Slack messages, reading Kubernetes resources, and retrieving metrics from an observability platform. After you configure and enable an MCP server in the console, agents decide on their own when an investigation needs an external capability. They call the tool directly, so you don’t have to copy data back and forth.
This page covers the “MCP” section inside the AI SRE console — that is, third-party MCP servers that agents call (you bring external tools in for agents to use).This is different from the Flashduty MCP Server in the developer docs: that is Flashduty’s own outbound MCP service. It connects Flashduty’s capabilities to third-party AI clients such as Claude or Cursor. One direction is “agents calling external services”; the other is “external clients calling Flashduty” — they are opposite flows, not the same thing.

What is MCP

MCP (Model Context Protocol) is an open protocol. It describes, in one uniform format, which tools a server exposes, what parameters each tool accepts, and what each returns. An agent can discover and call any MCP-compliant server without per-system adapter code. In AI SRE, MCP extends an agent’s capability from “built-in tools” to “any external system”:
  • You add an MCP server in the console (declaring its endpoint, transport, and authentication).
  • At session start, enabled servers visible in the current scope become available to the agent.
  • When needed, the agent discovers the tools that server provides and calls them directly.
MCP versus Skills: MCP provides the ability to connect external tools; a Skill provides the methodology for orchestrating those tools to accomplish a class of tasks. A Skill can declare which MCP server tool it needs inside SKILL.md using mcp:server/tool notation. See Skills for details.

Installing MCP servers from the marketplace

Go to Customize → Plugins → Overview to see Flashduty’s curated catalog of third-party MCP templates. Category tabs group the catalog: All / Observability / Cloud & Infrastructure / Database & Middleware / Code & Collaboration / Skill. Use the search box in the top-right corner to filter by name or description. The Recommended block above the catalog suggests up to 3 templates based on the alert integrations configured in your account (ranked by alert volume over the last 30 days). Templates that are already installed, need a Runner, or that you cannot install are not recommended.
1

Find the template to connect

Click a template row or its + to open the Connect dialog. Templates that need a self-hosted Runner when the account has none are marked Requires self-hosted Runner; templates you are not allowed to install cannot be clicked.
2

Review the template and fill in the connection parameters

The dialog shows the template’s vendor, official docs, description and example questions, followed by the connection parameters. Shared fields are used by the whole account; templates in member authorization (key) mode also ask for your own API token, which other members enter separately on first use.
3

Pick an execution environment and submit

Choose the execution environment to connect from (defaults to the session’s environment). For systems reachable only from an internal network, choose a self-hosted Runner on that network. For templates in member authorization (OAuth) mode, click the Authorize button and finish authorization in the browser; see Authentication.
4

Confirm the result

The system creates the connector and tests the connection. On success, you can click an example question to start a chat prefilled with it. On failure, the error is shown and you can Retry. Templates that need a Runner can also have AI SRE install missing command-line tools on the Runner (such as Node.js for npx or uv for uvx).
Connection parameters are fixed at the first submit; to change them later, edit the server under Plugins → MCP. If a connector with the same name already exists in the account, use it, or have an account owner or admin rename or delete it before connecting again.
MCP servers installed from the marketplace record their source template and are always Shared scope (available to all members of the account). Each template can be installed only once per account.
Templates that require a Runner (stdio servers started locally through uvx / npx) work only on a BYOC Runner you deploy; cloud Sandboxes do not support them. Make sure the account has a working BYOC Runner before installing. The catalog marks which templates need one. See Environments (BYOC).

Systems only Monitors can query

The catalog also lists systems queried through Monitors (such as Loki, VictoriaLogs and Oracle). They cannot be connected from the marketplace: once you add such a data source in Monitors, AI SRE can query it directly, without an address, token or Runner. Clicking one of these rows shows the account’s data sources of that type. Click Try it in chat if some exist, or Add more in Monitors → if none do (requires permission to add data sources). Some templates (such as Prometheus, MySQL, PostgreSQL, ClickHouse, Elasticsearch and SLS) support both a direct MCP connection and querying through Monitors. They are labeled MCP · Monitors, and you choose the access method when connecting.
Row names share a namespace with template names: if a system already has a template of the same name, only the template row is shown. On a private deployment without the Monitors module, these rows and the Monitors access method are not shown.

Adding an MCP server

Go to Customize → Plugins → MCP. The list page has three entries in the top-right corner:
  • Add from catalog: jumps to Customize → Plugins → Overview (the plugin catalog described above), where you pick and install a curated MCP template; it then appears in this list. It needs MCP create permission.
  • Add MCP: opens a form for defining an MCP server by hand (its fields are listed under “Basic Fields” below).
  • Add in chat: opens a session with a prefilled prompt. The agent asks whether you want to connect your own service or install an existing one, checks the plugin catalog first, and searches the web for candidates if nothing matches. The button appears only when the target scope is unambiguous (you have account-level permission, or exactly one of your teams is selected).
When the list is empty and you are not searching, the page shows “The plugin catalog has N built-in MCP servers. Install one to use it in sessions.” and, if you can create MCP servers, a Browse built-in MCP servers button that jumps there. When the catalog is empty or you are searching, the original “No MCP servers configured” prompt appears instead.

Basic fields

After an agent first lists a server’s tools, the system generates an AI description shown in the list, and keeps it updated as the tools change.

Transport

stdio is only available for BYOC environments (Runner deployed on your own machine). Cloud Sandboxes cannot launch local subprocesses. If you are using a cloud environment, use HTTP Streaming or SSE instead, and ensure the MCP server is reachable from the Sandbox network. For differences between environments, see Environments (BYOC).
stdio commands run on the machine where the Runner is deployed, not on your computer. Avoid personal paths such as /Users/<name>/.... Make sure the Runner has the required runtime installed, such as uv / uvx / node / npx, or the connection fails.
There is a default timeout for both connection and tool calls: 10 seconds for connection, 60 seconds for tool invocation.
Enable “Skip TLS certificate verification” only when the HTTPS server uses a self-signed certificate and the network is controlled. Never enable it for untrusted public endpoints.

Authentication

MCP servers support three authorization modes that determine how credentials are provided to the server:
All sessions call with the same service identity. You write credentials directly into the server configuration. For HTTP/SSE transport, put them in the Headers (JSON), for example { "Authorization": "Bearer xxx" }. For stdio transport, put them in the environment variables (JSON), for example { "API_KEY": "xxx" }. Suitable for internal systems accessed with account-level service tokens.Values under the following key names (case-insensitive) are encrypted at rest and shown masked. All other entries are stored as-is, so put credentials under such key names (for example Authorization, X-API-Key, GITHUB_TOKEN):
  • names containing KEY, SECRET, TOKEN, PASSWORD, PASSWD, CREDENTIAL or AUTHORIZATION;
  • names with a segment (split on non-alphanumeric characters) that is exactly PWD or PASS (such as REDIS_PWD, db-pass; BYPASS and COMPASS do not count).
Any session in the account can call the system with this identity. For personal data or high-risk write operations, use member authorization instead.
In member authorization modes, a tool call with missing credentials pauses and prompts the current member to enter a key or authorize, then continues.

MCP server authorization (settings)

Member authorization credentials (key or OAuth) are isolated per member. Besides authorizing when prompted in a chat, you can manage your own authorization from the MCP list; both paths save the same credential. Authorization status: the Member authorization column of the MCP list shows your own status:
  • Member authorization (OAuth): ● Connected / ○ Disconnected.
  • Member authorization (key): ● Saved / ○ Not set. The key is not checked for validity.
  • Both modes: ⚠ Expired (the OAuth token has expired and needs reauthorization).
Service authorization shows a dash (”—”). Click the authorization entry to open the Credential dialog; it manages only your own credential and does not change the server’s configuration. Actions in the Credential dialog:
  • Member authorization (OAuth): choose an execution environment (the cloud Sandbox or an online BYOC Runner. “Auto” is not available), then click Authorize or Reauthorize and finish in the browser window, with no tokens to copy. For OAuth services reachable only from an internal network, choose a Runner that can reach them. The last environment that authorized successfully and is still online is preselected.
  • Member authorization (key): click Enter Key or Update Key.
  • Revoke: deletes your saved credential.

Changing the destination leaves stored secrets behind

A save that changes the transport, the command, the args, or the endpoint (URL) counts as switching to a different connection destination. Args are compared as a list, so an unchanged list is not a change. Stored shared credentials do not go to the new destination: the protected values in the environment variables and Headers (key rules in Authentication above) must be re-entered as plaintext in the same edit. If you leave one out, send it empty, or send the mask back as-is, the save is refused and the form shows this error listing the affected key names:
The same edit also revokes every member’s credential for this server: member authorization (key) needs the key again, and member authorization (OAuth) needs a fresh authorization. The revocation happens before the save, so a save that ultimately fails can still cost members one more authorization. Replacing only the OAuth metadata (the value that names the token endpoint), through the API, does not require re-entering shared credentials, but it still revokes members’ credentials; emptying that value only triggers rediscovery from the unchanged endpoint and revokes nothing.

How protected values are shown

Protected values (environment variable and Header names that match those rules) always come back as the fixed mask ****, with no first or last characters revealed; sending **** back means “keep the stored value”. Members who can view but not edit the server see less of the connection config: args are not returned at all, and the endpoint (plus a proxy address set through the API) comes back with any user:password removed and every query-string value replaced by ****; an endpoint that cannot be parsed is left empty. Editors see the endpoint and args as entered.

Management & inspection

The MCP list shows each server you added with its name and description, scope (Shared or team), transport and enabled state. You can filter by All / Shared / Team and search by name, description, transport or URL. The built-in Flashduty MCP server is not listed.
Toggle the switch in the list. Only enabled servers are available to the agent; disabled servers are invisible to agents and cannot be called.A disabled connector cannot be used from the Connect entry and does not appear in the enabled section of Plugins → Overview. An account owner or admin must re-enable it under Plugins → MCP.
Click the edit button (or click the row) to open the form. You can modify the name, transport, description, endpoint/command, authorization mode, and scope. If you do not have edit permission, the form opens in read-only mode with an explanation.
Removes the MCP server from the current scope. Agents that depend on it will no longer be able to access its tools, and active sessions currently using it will fail. A confirmation prompt is shown before deletion.

Inspection: Viewing the tools a server exposes

The console does not list the tools an MCP server exposes, because the same server may be reachable with different tools in different environments.
To confirm which tools an MCP server exposes in a given environment, ask the agent in Chat to list that server’s tools. It shows the tools and how to use each one. See Conversations.
Reading Flashduty incidents, alerts and other data is a built-in agent capability (the Flashduty MCP server). It is available by default and is not configured here.

Cross-environment connectors

An MCP server may be bound to execution environments other than the current session’s. Because environments do not share a network, the session cannot call it directly, but the agent knows it exists:
  • The agent sees the connector’s name, purpose and available environments, and asking for it by name gets an explanation rather than “not found”.
  • The agent dispatches a subagent into the target environment to call it (see Agents) and brings the result back to the conversation.
  • Valid targets are account-level Runners, Runners of the team bound to the session, and the bound cloud environment. Any Runner that has ever connected is a candidate, including ones currently offline; Runners that have never connected are not. An offline Runner only reports that the environment is unavailable when the dispatch is attempted.
  • Only the main session can dispatch this way; subagents cannot dispatch further.
Conversely, to make a connector directly usable here, add the current environment to that server’s Execution Environment list (provided the server is reachable from this session). See Environments (BYOC) for how the environments themselves differ.

Scope

MCP shares the same two-level scope (shared / team) as other resources. For the general rules, see Scope and permissions. What is specific to MCP:
  • Servers installed from the marketplace are always shared, any member can install them, and they can’t be changed to team scope.
  • For a template that uses member authorization and also asks for a custom address, only the account owner or an admin can install it, so one member’s credentials are never sent to an address another member chose.

Skills

Skills call MCP-provided tools inside SKILL.md using the mcp:server/tool notation.

Agent

Agents share the same authorization modes and scope model as MCP.

BYOC

stdio MCP servers must run on a BYOC Runner; learn the difference between cloud Sandbox and self-hosted environments.

Console Chat

Observe how agents inspect and invoke MCP tools during a session.

Flashduty MCP Server (Developer)

The opposite direction: the official MCP service for connecting Flashduty into third-party AI clients.