MCP Connector
The MCP Connector links your agent to a remote Model Context Protocol (MCP) server. Once connected, your agent can automatically discover available tools and utilize them to complete requests.
Add one MCP connector for each MCP server. An agent can have up to 50 connectors.
Set Up a Connector
In Agent Studio, open Agent Tools, click Add Tool, and choose MCP Connector. Fill in:
- Name — a short name for the server, such as
GitHub. - Description — what the server is for, and when to use it.
- Auth header name (optional) — the header your server reads its credential from, usually
Authorization. Leave it empty if the server needs no authentication. - Server URL — the server's full
https://address. It often ends in/mcp. - Auth header value — the complete header value, for example
Bearer ghp_…. This field appears only when you set a header name. Once deployed, it's stored securely and shown as•••• Stored; use Replace to rotate it.
Test the Connection
Click Test connection under a server URL to confirm the MCP server is accessible. The test connects to the server the same way your agent does, and shows either the server's name and its MCP tools or the reason it failed. The test is optional, and you can deploy without it.
The test uses your stored credential only while the server URL and the auth header name still match what you deployed. Edit either one and the test runs with no credential, because a stored credential belongs to the server and the header it was saved for. Click Replace, enter the value again, and test it with the new settings.
Test and Live Connections
Each connector has a separate connection for each environment:
- Server Connection — Test is used when you deploy to Test.
- Server Connection — Live is used when you deploy to Live.
Point Test at a sandbox server or a test credential, and Live at the real one. The name, the description, and the auth header name are shared by both environments. When you deploy to Live, the confirmation shows the Live URL for each connector, and warns you if one is missing.
The Description Matters
Your agent decides whether to use a server from its name and description alone. Explain what the MCP server is for and when to use it.
A vague description such as "GitHub" makes your agent guess. It can then skip the server when it needs it, or connect when it doesn't.
There is no list of allowed tools. Your agent can use every MCP tool the server exposes, including tools the server adds later. Connect only servers you trust, and use a credential scoped to what your agent should be able to do. For example, use a read-only token if your agent only needs to read.
Removing a Connector
Press Remove on the card and confirm. Then deploy the agent build to Test and Live. The connector's settings and its stored credential are deleted.
What Is Supported
| Supported | Not supported |
|---|---|
| Remote servers over HTTPS (Streamable HTTP) | Local (stdio) servers |
One header of your choice for auth, e.g. Authorization: Bearer … | OAuth sign-in |
| No auth | Private-network or internal addresses |
| MCP tools | MCP resources and prompts |