Hello,
I’m trying to get the MCP server on RT 6.0.4 to work; I’ve been experimenting with a few different agents on my end, from the one that comes with mlx-serve to Lumo Desktop and yes, of course, Claude Code (macOS CLI).
However, even following the exact steps from Jim’s great walk-through video, and the announcement post here, I can’t get any agent to do anything other than confirm that there is an MCP server, and retrieve the list of tools.
Any attempted use of a tool results in a “fetch failed” or other similar (not very useful) error.
I am using a local LLM setup, and have configured Claude Code to point to my OpenAI compatible endpoint and set the appropriate environment variables. This works fine. I can interact with the agent, ask it to analyse code, do things, and everything works as it should.
Checking /mcp also shows RT, and 43 tools.
However, if I actually try to use one, such as from Jim’s video, it fails. Here’s an example:
❯ What user am I logged into RT as?
∴ The user is asking "What user am I logged into RT as?".
The available tools include an rt MCP server.
Specifically, mcp__rt__get_current_user seems like a perfect match for this question.
First, I need to load the tool schema for mcp__rt__get_current_user. Since it's a deferred tool, I should use ToolSearch to get its definition.
∴ The ToolSearch returned the tool definition. Now I can call it.
⏺ rt - get_current_user (MCP)
⎿ Error: fetch failed
As you can see, the “fetch failed” isn’t very useful.
I modified my Apache installation for RT to ensure that the SetEnv is present as:
SetEnvIf Authorization "(.*)" HTTP_AUTHORIZATION=$1
Which is set just after the ServerName in the declaration.
I’ve verified the token is correct with a simple curl command, and doing so also updated the “Last Used” timestamp in the RT Web interface for that auth token.
However, whenever I try to use Claude, or any other agent, it seems to get wedged. Eventually things time out, but, it goes in a long loop of arguing with itself about trying to make tool calls, check results, getting empty values, and so on.
I am probably missing something obvious here, but, I’ve verified the .mcp.json file repeatedly in the directory I’m in, know the token is right because I used it with the curl command, and the curl command was also able to do things like get a list of queues over the REST API, but the agent and MCP, nope.
The .mcp.json looks like so:
{
"mcpServers": {
"rt": {
"command": "npx",
"args": ["mcp-server-rt"],
"env": {
"RT_URL": "https://myhost:9009",
"RT_TOKEN": "1-99-longstuffhere"
}
}
}
}
Yes, I’ve modified it to remove my token. My RT instance runs on a non-standard port, and has a non-globally-trusted TLS certificate. However, I would have assumed if the agent cared about that, it wouldn’t have been able to get the list of 43 tools to start with?
I installed Node.js from homebrew, which is currently v26.13.1, and npx is 11.20.0.
Thank you for any pointers!