Dashboards For Claude Agent Work

Dashboard The Work, Not The Hype

A Claude dashboard should answer operational questions: Are sessions completing? Which tools are slow or failing? Are permissions being denied? Are token and cost patterns changing? Are MCP calls attributed to the right user? Are background agents getting stuck?

Avoid vanity charts that count prompts without outcome. A session count only matters when it can be tied to success, failure, cost, or risk.

Core Panels

Start with a small dashboard that can survive production review. Show sessions started, successful and failed outcomes, estimated cost, token usage, API latency, tool execution time, tool decision counts, and top failing tools. Add dimensions for repository, team, model, query source, and user identity where policy allows.

Add a separate security view for permission mode changes, hook blocks, MCP server connections, detailed MCP tool calls where enabled, and unusually broad file or shell activity.

• Session health: starts, stops, stop failures, background restarts.
• Efficiency: tokens, cost estimates, cache behavior, turns per completed task.
• Tooling: tool latency, failures, denials, Bash command categories.
• Security: permission changes, MCP detail, hook blocks, identity attribution.

Alert Ideas That Are Worth Paging

Most agent alerts should go to a workflow channel, not a pager. Page only when agent activity can affect production, customer data, credentials, or high-cost automation. For everything else, create daily or per-PR review queues.

Good alerts are tied to action. An alert that says tokens are high is weak. An alert that says a production-bound workflow exhausted retries, used bypass permissions, and modified deployment files is actionable.

Correlate Back To Evidence

Every dashboard row should lead back to evidence: session ID, transcript path or archive key, trace ID, PR link, commit SHA, run ID, and user identity where available. Without those links, a dashboard becomes a wall of symptoms.

The transcript and dashboard should disagree gracefully. If telemetry says a tool failed but the transcript summary says work finished, send the reviewer to the raw event and final repository state.