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

Rius lets you query your agents’ traces from Claude Code by adding the Rius MCP server and asking questions in plain language. You add it with one command, sign in with your Rius account in the browser the first time a session uses it, and Claude Code answers from your real telemetry. If you installed the Rius plugin for Claude Code, the server is already registered as rius.

How it works

The Rius MCP server is a remote endpoint at https://mcp.eu.console.rius-glassflow.com/mcp, reached over streamable HTTP under the server key rius. Claude Code reads the tools the server offers and picks the right ones for your question. You do not name tools yourself.

Three points shape how it behaves:

• Querying is a second sign-in. It is separate from the sign-in that sends traces. Tracing your own sessions with the plugin does not by itself authorize queries, and the reverse is also true.
• An interactive session acts as you. The tools see the workspaces your user can access and act with your role. Every member can read and acknowledge a firing alert. Creating API keys and managing alert rules or channels needs an admin, as in the console.
• Tools declare whether they read or write. Each tool carries a hint, so a client that honors hints can auto-approve reads and ask you to confirm writes.

This is the server your client queries. It is not the SDK’s mcp integration, which traces your own agent’s outgoing MCP tool calls.

Set it up

1. Check whether you already have it. If you installed the Rius plugin for Claude Code, it registers the server as rius and you can skip to step 3. The Coding agents page covers the plugin and what it traces.

2. Add the server. Run this in your terminal:

   claude mcp add --transport http rius https://mcp.eu.console.rius-glassflow.com/mcp

3. Sign in. Start a Claude Code session and ask a question about your traces. The OAuth sign-in opens in your browser the first time a session uses the server, and uses your Rius account. If the client shows the server as connected but lists no tools, finish the browser sign-in. If the sign-in loops, remove the server and add it again, because a stale cached token is the usual cause.

4. Verify. Ask:

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

   Claude Code should call the get_me tool and answer with your email and a table of workspaces.

5. For a session with no browser, use a read-scoped key. CI jobs, cron schedules and autonomous coding sessions send an API key instead of using OAuth. Mint a key with the read capability from the Connect an agent button on the console’s MCP page, store it as the RIUS_API_KEY environment variable, then run the command below. It uses the name rius-key, so it sits next to the plugin’s OAuth rius server.

   claude mcp add --transport http rius-key https://mcp.eu.console.rius-glassflow.com/mcp \
     --header "Authorization: Bearer ${RIUS_API_KEY}"

   Ingest keys, the kind the SDK uses, cannot query. They return a 403 for a missing scope. A key is bound to the one workspace it was created in.

The MCP page has the snippets for other clients and the full tool reference.

What you get

You ask in plain language and Claude Code answers from your own telemetry, without leaving the terminal.

• Traces. Ask “List the failed runs from this week and show me what went wrong in the worst one.” Claude Code lists the failed traces, then opens the span waterfall of the worst one. Tools such as list_agent_traces, get_agent_trace and list_spans cover this.
• Traffic and cost. Ask “Summarize my agent traffic over the last 24 hours.” You get traces, LLM calls, error rate, cost, tokens and latency percentiles. Ask “Which service is spending the most right now?” to find the biggest spender. Costs are list-price estimates in US dollars, not billed amounts.
• Agents and versions. Ask about error groups, lifecycle events or a comparison of two agent versions with tools like agent_error_groups and compare_agent_versions.
• Alerts. Ask what is firing with list_alerts and get_alert, check swallowed errors with list_silent_failures, or read rules and channels with list_alert_rules and list_alert_channels.

A read-only key also reaches the alert reads. A key minted with the alert-management scopes can change alert rules or channels in its workspace. A key never acknowledges a firing, because that stays a person’s action.

When something else fits

• If you want to browse a waterfall visually or share a filtered view, the Rius console is the better place, and the MCP server cannot create, rename or delete workspaces or change members either.
• If you work in another client such as Codex, Cursor or Claude Desktop, the MCP page has the matching connect steps.
• For a CI job or scheduled agent, use the read-scoped key in step 5 instead of the browser sign-in.

Related