Tool-Usage Analytics
Per-deployment tool call volume, error rates, latency, and invocation history.
Tool-usage analytics answers: which MCP tools is this server calling, how often, how fast, and how often do they fail? — plus a searchable audit log of individual invocations.
Analytics are per deployment. Open a deployment in the dashboard and use the Analytics tab.
What you can see
- Call volume — total tool calls over a time window (24h, 7d, 30d, or custom)
- Error rates — failed calls vs total, overall and per tool
- Latency — p50 and p95 tool-execution duration (milliseconds)
- Per-tool breakdown — leaderboard table and latency bars by tool name
- Invocation log — paginated, newest-first list of individual tool calls (expand a row for the raw audit payload)
Honest limits (v1)
Keep these in mind when interpreting charts and planning workflows:
| Limit | Detail |
|---|---|
| Latency is tool execution only | Measured at the ToolHive proxy — how long the MCP server took to handle the call. It is not LLM latency or full client end-to-end time (the model runs in the user’s client, which the platform never sees). Charts are labeled as tool execution time. |
| No token or cost data | Out of scope for v1. Do not expect token counts or dollar cost on this page. |
| Near-real-time, not live | Invocations arrive via a log pipeline with a few seconds of lag. Refresh or poll; there is no websocket/SSE live feed yet. |
| One deployment at a time | No project- or account-level rollups in v1. The sidebar Analytics item (account-wide) is separate and marked Coming Soon. |
| No historical backfill | Data starts when analytics collection is enabled for traffic. Empty state is normal for brand-new or quiet deployments. |
Where in the UI
- Open Deployments in the sidebar.
- Select a deployment.
- Open the Analytics tab.
- Pick a time range and refresh as needed.
Empty state: “No tool calls recorded yet” means either no traffic yet, or no data in the selected window — not a broken page.
API
Same Supabase session bearer auth as the rest of the dashboard API. Base URL: https://api.mcplambda.io (prod). Requests for a deployment you do not own return 404.
Tool usage (charts / KPIs)
GET /v1/deployments/{id}/analytics/tool-usage?from=&to=&bucket=| Query param | Default | Notes |
|---|---|---|
from | now − 7 days | RFC3339, inclusive |
to | now | RFC3339, exclusive |
bucket | day | minute | hour | day | week |
Response body (no data envelope) includes:
series— rows per(bucket_start, tool_name)withtotal_calls,error_calls,p50_duration_ms,p95_duration_mssummary— per-tool aggregates over the whole window (use this for KPIs and the tools table)
series / summary may be null or [] when there is no data.
Invocation audit log
GET /v1/deployments/{id}/analytics/invocations?page=&page_size=page is 1-based (default 1); page_size defaults to 20. Newest first. Each row includes tool name, outcome (success | error), duration, optional caller identity fields, and a raw payload for detail views.
Operational tips
- Prefer p50 and p95 together — never rely on a bare average.
- High error rate on one tool often points at missing secrets, upstream API errors, or auth misconfiguration on that tool path.
- Use the invocation log when a single failed call needs the raw ToolHive audit record.