How do I query my agents’ traces from inside Claude Desktop?

Rius connects to Claude Desktop through its MCP server, so you can ask about your agents’ traces, cost and alerts in plain language. You add the server to claude_desktop_config.json, restart Claude Desktop, and sign in with your Rius account in the browser. Claude then answers from your real telemetry: trace lists, cost and token summaries, span waterfalls and alerts.

How it works

The Rius MCP server is an endpoint that Claude Desktop connects to. It is separate from tracing: you send traces with the SDK or a plugin, and you query them through this server.

• Transport and URL: streamable HTTP at https://mcp.eu.console.rius-glassflow.com/mcp. The server key is rius.
• Sign-in: your Rius account over OAuth 2.1. Claude Desktop runs the browser flow for you, and the server sees the workspaces your user can access.
• Tools: you ask a question and Claude picks the right tools. For traces that includes list_agent_traces, get_agent_trace, list_spans and agent_traces_summary. For cost and traffic it includes workspace_metrics_overview. For alerts it includes list_alerts and get_alert.
• Workspaces: if you belong to several organizations, get_me lists them all with their workspaces. When the same workspace name exists in more than one organization, pass it as org/workspace.

If you use Claude Code, Codex (beta) or Cursor (beta) with the Rius plugin, that plugin already registers this server as rius. See Coding agents for those setups. Claude Desktop needs its own entry, as below.

Set it up

1. In Claude Desktop, open Settings, then Developer, then Edit Config. Add the Rius server to claude_desktop_config.json:

{
  "mcpServers": {
    "rius": {
      "url": "https://mcp.eu.console.rius-glassflow.com/mcp"
    }
  }
}

2. Restart Claude Desktop so it loads the new server.

3. Complete the browser sign-in with your Rius account. Most clients show the server as connected before the sign-in has finished, so confirm you completed it if no tools appear.

4. Verify the connection by asking:

   Who am I in Rius, and which workspaces can I access?

   Claude should call get_me and answer with your email and a workspace table. If it does, everything works.

5. If the browser sign-in loops or Claude reports an auth error, remove and re-add the server. A stale cached token after a long idle period is the usual cause.

If you want a connection without the browser flow, use a read-scoped API key. Mint it from the Connect an agent button on the console’s MCP page, or in Settings → API keys with the read capability. Then add a headers block to the same entry:

{
  "mcpServers": {
    "rius": {
      "url": "https://mcp.eu.console.rius-glassflow.com/mcp",
      "headers": { "Authorization": "Bearer ri_..." }
    }
  }
}

The key is shown once, so store it as a secret. It reaches only the workspace it was created in. Ingest keys, the kind the SDK uses, cannot query through the server. The MCP page has the full list of what a key can reach and the connection snippets for other clients.

What you get

Once connected, you ask questions in Claude Desktop and get answers drawn from your workspace.

• Traces: “List the failed runs from this week and show me what went wrong in the worst one.” Claude filters the trace list by status, then opens the span waterfall of the trace it picks.
• Traffic and cost: “Summarize my agent traffic over the last 24 hours.” You get traces, LLM calls, error rate, cost, tokens and latency percentiles. “Which service is spending the most right now?” works the same way. Costs are list-price estimates in US dollars.
• Agents and users: ask which agents exist, how their versions compare, which errors repeat, and which users were affected.
• Alerts: ask what is firing, read the details of one alert, and look at silent failures. With an OAuth sign-in, any member can acknowledge a firing alert. Creating or changing alert rules and channels needs an admin role, the same as in the console. An API key never acknowledges a firing.

Tools carry hints that say whether they only read or change something. Clients that honor the hints can auto-approve reads and ask you to confirm writes. Rius logs each MCP call for 30 days, with structural arguments only. It never stores tool results or free-text search terms.

When something else fits

• If an agent runs without a browser, such as a CI job or a scheduled run, use the read-scoped key path instead of the interactive sign-in.
• Creating, renaming or deleting workspaces and changing members happen in the console, not through the MCP server.
• A key cannot mint other keys or read spend notices. Use an interactive session or the console for those.

Related