Skip to main content
The MCP page connects Model Context Protocol (MCP) servers to Arklex Platform. Once a server is connected and an Owner approves its read-only tools, custom metrics can call them during evaluations, and the scenario assistant can use them as a source. For example, the judge can look up an order’s real status and check whether the agent reported it correctly.

Add a server

Go to MCP in the sidebar and click Add Server. Fill in the Add MCP Server dialog:
string
required
A name for the server, for example “Orders API (production)”. Names must be unique in your organization.
string
required
Streamable HTTP, the only supported transport. Servers that offer only stdio or the older HTTP+SSE transport can’t be connected.
string
required
The server’s MCP endpoint, which often ends in /mcp, for example https://example.com/mcp. On cloud, it must be reachable from the public internet, because private and localhost addresses are blocked. Self-hosted deployments allow them by default. See the note under Endpoint URL.
object
One header name and value for authentication. The name defaults to Authorization. If your server doesn’t accept Authorization, use the header it expects, such as X-API-Key. The value is encrypted at rest and shows as **** after you save.
Click Test connection to check the server and list its tools, then click Add and Connect. You can’t edit a server after you add it. To change its URL or credentials, delete the server and add it again.

Read the server list

The summary at the top shows Reachable servers and Enabled tools. The table lists each server:

Manage tools

Click a server to open its details. The Tools tab lists every tool the server offers. Click Re-run tools/list to refresh the list. For each tool, you can see:
  • An Enabled switch. Only enabled tools can be used.
  • A safety label: Read-only, Write-capable, or Unknown safety.
  • A Judges and scenario generation may call switch, which approves the tool for evaluation judges and the scenario assistant. Only Owners can turn it on, and only for Read-only tools, so neither can change data in your systems.
Tools that the server marks as read-only are enabled automatically. If a tool’s definition changes on the server, Arklex removes its approval, and an Owner must approve it again.

Test a tool

Owners can run a tool by hand to check what it returns. Click Test this tool, fill in the Arguments (or click Edit as JSON), and click Run tool. The server and the tool must both be enabled.
Run tool really calls the tool, even a Write-capable one, so it can change data in your systems.

Use tools in evaluations and scenarios

To use an approved tool in evaluations, add it to a custom metric. See Check facts with MCP tools. To build scenarios from live records, pick approved tools as a source when the assistant asks which knowledge to use. See Build with the assistant. During the evaluation, the judge calls the tool and scores the agent against the results. In the conversation view, you can see the Evidence used for this score. When you re-evaluate, choose whether to reuse the saved tool results or fetch new ones. See Re-evaluate.

Roles

Limits

  • Connection tests time out after 10 seconds, and tool calls time out after 30 seconds.
  • Each metric can make up to 5 tool calls per turn and 20 per conversation on turn scope, or 5 in total on conversation scope.
  • Arklex lists up to 500 tools per server.
  • On cloud, server URLs must be public.