Skip to content
SentientOne

MCP tools

Connect an MCP server to an agent

Add a remote MCP server on the agent's Tools tab, test the connection, discover its tools and try them in the Playground.

Guide2 min readUpdated 23 October 2026

This guide adds a remote MCP server to an agent and checks the agent can call its tools.

Before you start

  • An MCP server reachable on a public internet address over https:// (or http://), speaking the streamable HTTP transport: JSON-RPC over POST, answering with JSON or an SSE-formatted response.
  • Its URL, usually ending in /mcp, and any credentials it needs. See Authentication.
  • An agent whose model supports tool calling.

Steps

  1. 1

    Open the Tools tab

    Open the agent and go to Tools, then Add server.

  2. 2

    Name it and pick the transport

    Enter a Name and set Transport to HTTP (streamable). Leave Status on.

  3. 3

    Enter the URL and headers

    Paste the server's URL, for example https://mcp.example.com/mcp. Add any Custom headers as one KEY=VALUE per line.

  4. 4

    Set authentication

    Choose None, Bearer token, API key header or Username and password, and fill in the secret. It is encrypted when saved and never shown again in full.

  5. 5

    Add and test

    Click Add server, then Test connection on its row. The test runs the MCP handshake and lists the tools. If the URL has no path, the test also tries it with /mcp added.

  6. 6

    Refresh tools when the server changes

    SentientOne stores the tool list it discovered and uses that list at chat time, so a reply never waits on discovery. After you add or change tools on your server, click Refresh tools.

  7. 7

    Try it in the Playground

    Ask a question that needs the tool. The reply shows the tool calls beside it; open the details panel's Tools tab to see each call's parameters.

You're done when

The server's row shows Connected with its tool count, and a Playground reply shows a tool call with the arguments you expect and an answer built from the result.

About the transport options

Option on the Tools tabWhat happens
HTTP (streamable)Supported. One POST per JSON-RPC call (initialize, tools/list, tools/call), with Accept: application/json, text/event-stream; JSON and SSE-formatted responses are both read. The Mcp-Session-Id the server returns is reused.
SSE (server-sent events)Treated exactly like HTTP: the same POST requests. Servers that only offer the older two-endpoint SSE transport (a GET stream plus a separate message endpoint) won't work.
stdio (local process)Not supported in the cloud. Servers added this way show as Local; they aren't discovered and their tools can't be called.

If something goes wrong

"MCP servers must be on a public internet address"

The URL points at localhost, a .local or .internal name, or a private, loopback, link-local or carrier-grade NAT address. Expose the server publicly (for example behind your API gateway) and use that URL. See Network rules.

The test fails with 401 or 403

Check the authentication type and the secret. For an API key header, the header name must be exactly what your server expects.

The test passes but the agent never calls the tool

Click Refresh tools so the stored tool list is current, make sure the server's Status is on, and give tools clear descriptions: the model decides from the name and description. More in MCP server not connecting.

Related

Still stuck?

Email us with what you tried and any request id. A person replies.

support@sentientone.ai