Inspiration
Media-rendering incidents are costly because operators must jump between logs, traces, dashboards, and runbooks before they can decide what to do. CutCue turns that fragmented investigation into a concise, evidence-first incident brief.
What it does
CutCue is an incident-response copilot for media-rendering teams. It uses Gemini to reason over live Grafana Loki telemetry through Grafana's official MCP server, correlate HTTP 503 failures with trace IDs, explain the likely impact, and recommend the next safe action. Operational changes remain behind explicit human approval.
How we built it
The agent is implemented in Python with Google ADK and Gemini. A dedicated read-only deployment of Grafana's official MCP server runs on Google Cloud Run and accesses Grafana Cloud through a narrowly scoped token stored in Google Secret Manager. The repository includes the agent, deterministic demo telemetry, Grafana dashboard provisioning, deployment configuration, tests, and a public demonstration page.
Challenges
The main challenge was preserving a trustworthy evidence chain while keeping credentials and write operations out of the model's reach. We separated read-only investigation from approval-gated actions, constrained the MCP tool surface, and made the demo reproducible with traceable synthetic incidents.
Accomplishments
CutCue correlates real dashboard evidence, log events, and trace identifiers into a reviewable incident brief; uses the official Grafana MCP integration; deploys with least-privilege secret handling; and provides a public repository, live demo page, dashboard, and concise video walkthrough.
What we learned
Agentic operations are most useful when every conclusion can be traced back to observable evidence and every consequential action remains under human control.
What's next
Next steps include richer trace correlation, configurable runbooks, approval workflows for remediation, and evaluation against real production incident datasets.
Log in or sign up for Devpost to join the conversation.