Cursor's own usage console tells you what your team spent. This dashboard tells you what the agent did to spend it: which tools it ran, how long each took, which ones failed, and how long a turn takes end to end. It covers token spend by model, turn and tool latency, the mix of tools the agent reaches for, and tool failure rate with the individual failures behind it.
Panels read the OTLP traces produced by Cursor's agent hooks, so nothing beyond the setup in the guide is required. Queries filter on service.name = 'ide-agent' and on gen_ai.client.hook.event, which is what separates the events a turn emits: afterAgentResponse carries token counts, PostToolUse carries completed tool calls, and PostToolUseFailure carries failures. Agent, Repository, and Model variables scope every panel, so one dashboard covers as many repositories and models as your team works across.
Dashboard Preview
Recommended. Uses the V2 dashboard schema and needs SigNoz v0.135.0 or newer.
Import it in SigNoz with Dashboards → + New dashboard → Import JSON. Import guide
What This Dashboard Monitors
This dashboard tracks the telemetry Cursor's agent hooks emit to help you:
- Monitor Token Spend: Track input and output tokens per turn to understand workload and cost
- Compare Model Usage: See which models carry the workload, grouped at span level because Cursor switches models mid-session
- See What the Agent Did: Get the per-turn tool sequence that the vendor console does not show
- Measure Turn Latency: Watch p50, p95, and p99 of a complete agent turn rather than a single model call
- Find Slow Tools: Break latency down per tool to find the step that stalls a turn
- Track Reliability: Watch tool failure rate and read the individual failures behind it
- Attribute Work: Break spend down per repository and branch
Panels Included
The Overview section summarizes activity in the selected time range:
| Panel | Description |
|---|---|
| Input Tokens | Total input tokens across every turn. Input drives spend here, not output. |
| Output Tokens | Total generated tokens, typically one to two orders of magnitude below input. |
| Agent Turns | Count of prompts answered, one per gen_ai.client.generation span. |
| Tool Failure Rate | Failed tool calls over all completed tool calls. Turns red above 10 percent. |
Tokens
- Input Tokens Over Time: Input tokens per interval. Sustained growth without a matching rise in turns points at context bloat rather than more work.
- Output Tokens Over Time: Output tokens per interval. Charted separately because a shared axis with input flattens output to nothing.
- Tokens by Model: Turns and token split per model, grouped at span level so a session that switches models is attributed correctly.
Latency
- Turn Duration: p50, p95, and p99 of a complete turn. A widening p99 against a flat p50 usually means a few very long agent loops.
- Tool Latency p95 by Tool: p95 per tool. Shell sitting above the rest is the normal shape, since it waits on real commands.
Tools
- Tool Call Distribution: Share of completed calls per tool. Search and read tools dominating is the normal shape for a coding agent.
- Tool Calls Over Time: Completed calls per interval by tool, useful for spotting a window where one tool suddenly dominates.
- Tool Performance: Call volume with average and p95 latency per tool. A low average with a high p95 means the tool occasionally blocks the agent.
Errors
- Tool Failures Over Time: Failures by tool. A sustained line for one tool is usually a broken path rather than transient errors.
- Recent Tool Failures: The individual failed calls behind the rate, with tool, model, duration, and repository.
