Integrations
MCP servers
Connect any MCP-compatible tool so the assistant can use it.
MCP — the Model Context Protocol — is a standard way for an AI assistant to use an external tool. If a service publishes an MCP server, connecting it here teaches the assistant to use that service.
This is the advanced corner of the product. You do not need it to run servers.
What it gives you
Once connected, the tools that MCP server offers become available in chat. With GitHub connected, for example, you can ask about repositories, issues, pull requests and Actions runs, and the assistant reads them directly.
Adding one
MCP → Add. There are four built-in guides — GitHub, Jenkins, Grafana and any other MCP server — and picking one fills in most of the form and shows exactly what to create on the other side.
Connection type
| Transport | Use when |
|---|---|
| HTTP | A hosted MCP endpoint. The usual choice |
| SSE | Server-sent events, for servers that use it |
| stdio | A command run on the machine DevOps Agent runs on |
For HTTP / SSE you give the Server URL and any headers — typically
Authorization: Bearer <token>, one per line.
For stdio you give a command (e.g. npx), arguments one per line, an
optional working directory and environment variables as NAME=value,
one per line.
Example: GitHub
- Transport: HTTP
- URL:
https://api.githubcopilot.com/mcp/ - Header:
Authorization: Bearer <your token>
Create the token at GitHub → Settings → Developer settings → Personal access tokens. A fine-grained token limited to the repositories the assistant should see is the right choice; typical permissions are Contents (read), Issues, Pull requests and Metadata, plus Actions if you want workflow status. GitHub shows the token once.
Safety — read this before connecting one
An MCP server is code someone else wrote, given access to your chats.
- Only connect servers you trust. Prefer official endpoints from the vendor whose service it is.
- Scope the token tightly. Repos it needs, not your whole account.
- The approval gate covers MCP tools — a tool that changes something needs Execute mode and your Yes.
- But: an MCP server declares which of its own tools are read-only, and we trust that declaration. A badly written or malicious server could label a writing tool as read-only. This is the reason for the first bullet.
Troubleshooting
| Problem | Check |
|---|---|
| Tools do not appear in chat | The server URL and the token. Re-save to re-test |
| 401 / 403 | Token expired or lacking permission |
| Works sometimes | Rate limits at the other end |
| Assistant ignores the tools | Mention the service by name in your message — the tools it loads depend on what you ask about |
Permissions
Adding, editing and removing MCP servers: Owner or Admin. Using read-only MCP tools: everyone. Using MCP tools that change things: Operator and above, in Execute mode. See Roles.