Skip to main content
The Insights page provides organization administrators with detailed, real-time observability into LLM usage, token consumption, and financial spend. It helps you monitor costs, track performance, and drill down into usage at the workspace, agent, provider, and model level.

​
Accessing the dashboard

The Usage dashboard is admin-only. To access it:
1

Open your organization dashboard

Navigate to your organization dashboard.
2

Click Insights

In the left navigation sidebar, click Insights (above Settings).
Access to the Insights tab is restricted to users with the Admin role in the organization.

​
KPI strip

At the top of the dashboard, the KPI strip displays high-level metrics for the selected time window:
  • Spend: Total estimated spend in USD, based on standard list prices.
  • Tokens: Total token volume (Prompt + Output + Cache Write). Prompt tokens are cache-read inclusive, so they already contain Cache Read tokens.
  • Sessions: Distinct user/agent sessions that recorded usage in the selected period.
  • Retries: Total number of retried attempts across the group’s LLM calls (excludes the first attempt).
  • Cache Rate: The percentage of prompt tokens served from cache (e.g., Anthropic Prompt Caching or OpenAI Cached Input).
  • Output Ratio: Output tokens as a share of total billable (uncached) input plus output tokens. Dividing by billable input prevents cache replays from diluting the ratio and hiding over-generation.
  • VS Previous Period: Percent change in spend compared to the previous equal-length time window (e.g., comparing the last 30 days to the 30 days before that). If no baseline data exists, this displays “No prior data”.

​
Usage over time

The Usage Over Time chart displays daily metrics in an interactive bar chart.
  • You can toggle between Tokens (daily token volume) and Spend (daily USD spend) using the segmented control.
  • Hovering over any bar on the chart displays a tooltip with the formatted daily value and date.

​
Today (hourly) window

Select the 1d (“Today (hourly)”) option in the time window range selector to switch the chart into an hourly view. This view renders today’s 24 hourly bars, resolved in your local timezone from browser detection.
The hourly (1d) view is spend-only. The Tokens toggle is unavailable while the 1d window is active.

​
Model-stacked spend and interactive legend

In the Spend view, the chart stacks each bar by model so you can see how much each model contributed to spend in a given day or hour. An interactive model legend appears below the chart:
  • Click a model in the legend to isolate that model’s series.
  • Click the same model again to toggle it back and restore the full stacked view.

​
Bar drill-down

Click any daily or hourly bar to reveal an attribution drill-down panel below the chart. The panel’s contents depend on the active time window:
  • Day drill: Lists the agents that spent on that day. Each agent shows its per-agent spend total and visual model dots indicating which models it used.
  • Hour drill: Lists both the agents and the triggers (webhook or time) that incurred spend during that hour. Each entry shows its spend total and model dots.
Click an agent or trigger in the panel to jump directly to its row in the usage breakdown. You can narrow down the metrics shown across the entire page using the filters in the header:
  • Time Window: Choose between 1d (“Today (hourly)”), 7d (last 7 days), 30d (last 30 days), or 90d (last 90 days).
  • Provider: Filter metrics to a specific LLM provider (e.g., Anthropic, OpenAI, Gemini).
  • Model: Filter metrics to a specific model (e.g., claude-sonnet-4-6, gpt-4o).
Active filters are displayed as removable chips below the header. Clicking the X on a chip clears that specific filter, or click “Clear filters” to remove all active filters.

​
Usage breakdown

Below the chart, the breakdown panel splits organization spend and token counts five ways across tabs:

​
Workspaces

Shows spend, token count, session count, and retry count (num_retries) for each active workspace. Clicking on any workspace row drills down into its specific usage details.

​
Agents

Shows spend, token count, session count, and retry count (num_retries) attributed to specific versioned agents. Clicking on an agent row drills down into its specific usage details.
Usage from tool tasks or direct assistant tasks without a committed agent version is omitted from the agent breakdown but is captured in the organization/workspace totals.

​
Users

Available for organizations. Shows spend, token count, session count, and retry count for each active user in the organization. Clicking on a user row drills down into their specific usage details.

​
Providers

Shows spend and total tokens grouped by inference provider.

​
Models

Shows spend and token counts grouped by LLM model name.
  • Rates & spend breakdown: Click any model row to expand it in-place and reveal the per-million-token list prices (Input, Output, Cache Read, Cache Write) and the exact calculated spend contribution for each token type.
  • Estimated pricing indicator: Models that Guild priced using the default list price display a warning indicator next to the model name. Hover the indicator to see a tooltip, or expand the row to read the estimate note in the rate panel.

​
Drilling down (workspace, agent, and user scopes)

Clicking on a row in the Workspaces, Agents, or Users breakdown navigates to a dedicated drill-in view:
  • The breadcrumb navigation changes to Usage / Workspace · <Name>, Usage / Agent · <Name>, or Usage / User · <Name>.
  • A prominent scope pill badge is displayed below the header indicating the current drill-in filter.
  • The KPI cards, chart, and remaining breakdown tabs are automatically scoped to represent only the selected workspace, agent, or user.

​
User drill-in

Clicking a row in the Users breakdown scopes the metrics, charts, and breakdowns to reflect the selected user’s spend. User attribution is based on the chat-session initiator, representing interactive spend. Automated trigger and test spend is not user-attributed, so the Users and Triggers tabs are hidden under a user scope. A View user profile header action link opens the user’s Agent Hub profile.

​
Efficiency findings

The findings engine proactively scans LLM usage and surfaces actionable recommendations to reduce wasted spend. Each waste-detection check flags a specific pattern so you can act on it:
  • Prompt cache written but rarely reused (cache_write_without_read): Triggers when cache-write volume is high but the read-to-write reuse ratio is extremely low, meaning you pay to write prompt cache that is barely read back.
  • Output-heavy generation (over_generation): Triggers when output tokens dominate the billable input plus output mix, indicating verbose or costly output.
  • High LLM retry rates (retry_burn): Triggers when LLM calls retry frequently, indicating potential performance or reliability issues that burn spend and latency.