exploring-mcp-tool-usage
Point de départ pour explorer comment les outils d'un serveur MCP PostHog sont utilisés — achemine une question large vers l'outil typé qui y répond. À utiliser lorsque l'utilisateur demande « comment…
npx skills add https://github.com/posthog/ai-plugin --skill exploring-mcp-tool-usageExploring MCP tool usage
Any MCP server instrumented with the @posthog/mcp SDK emits a $mcp_tool_call
event every time an agent invokes a tool. This skill is the front door for a
user who knows they want to look at their MCP tool usage but hasn't picked a
specific question. Offer the menu below, then route to the tool — or the focused
skill — that answers what they choose.
Governed metric first
When the question asks for an MCP failure rate or percentage, call posthog:metric-list before any typed tool or SQL recipe and look for mcp_tool_call_fail_pct. Run an approved, non-drifted match with posthog:data-catalog-metric-run for the canonical headline. If the user also asks which tool or harness drives failures, answer the headline first, then use the per-tool workflows below for a noncanonical breakdown. If no governed metric matches, say so and label the derived rate noncanonical.
Every per-tool tool here takes a
toolName (the effective tool name — resolved server-side, so pass the name the
agent actually invokes — except posthog:query-mcp-tool-failures, which
matches $exception events and so takes the raw registered $mcp_tool_name)
plus a dateRange, and runs the same query runner the tool-detail UI uses. So
results match the UI, and you never hand-write the HogQL.
Suggested questions
Lead with these when the user is unsure what to ask:
| Ask the user… | Answered by |
|---|---|
| "Which tools fail most, or are slowest?" | exploring-mcp-tool-quality (ranks all tools), then posthog:query-mcp-tool-stats to drill in |
| "How is tool X doing overall?" | posthog:query-mcp-tool-stats — calls, errors, p50/p95, users, sessions, intents |
| "How has tool X trended?" | posthog:query-mcp-tool-daily-stats — day-by-day series |
| "Why is tool X failing?" | posthog:query-mcp-tool-failures — top error messages, by harness (raw tool name) |
| "Who uses tool X the most?" | posthog:query-mcp-tool-top-users — top callers (incl. person email/name) |
| "What gets called right before/after tool X?" | posthog:query-mcp-tool-neighbors (neighborDirection: before/after) |
| "What are agents trying to do with tool X?" | posthog:query-mcp-tool-sample-intents — recent agent intents |
| "What description is tool X registered with?" | posthog:query-mcp-tool-descriptions — distinct descriptions seen |
| "Which harnesses use my MCP, how reliably?" | posthog:query-mcp-harness-breakdown — calls/errors/sessions per client |
| "What are agents trying to do, across all tools?" | exploring-mcp-intent-clusters — semantic goal clusters |
| "Who is connecting, and how active are they?" | posthog:mcp-analytics-sessions-list — one row per session, with client and person |
| "What did this one session do?" | exploring-mcp-sessions — a single agent run's tool sequence |
Finding the tool name
The per-tool tools need a toolName. If the user named a tool, pass it. If they
asked a broad "which tool…" question, start with exploring-mcp-tool-quality to
rank the tools, pick the one that stands out, then drill in with the per-tool
tools above. The name to pass is the effective tool name (the inner tool for
single-exec wrapper calls) — the same string the tool-quality ranking returns.
The one exception is posthog:query-mcp-tool-failures, which matches $exception
events by the raw registered $mcp_tool_name, not the effective inner tool.
How to use a per-tool tool
Call it with the tool name and a window, e.g. for the headline numbers of a tool:
posthog:query-mcp-tool-stats { "toolName": "<tool>", "dateRange": { "date_from": "-7d" } }
Then offer a natural follow-up from the menu — e.g. after
posthog:query-mcp-tool-stats shows a high error rate, reach for
posthog:query-mcp-tool-failures; after it shows broad reach, reach for
posthog:query-mcp-tool-top-users or posthog:query-mcp-tool-neighbors.
When to drop to SQL
Covered by a typed tool — don't hand-write SQL for these:
| Question | Tool |
|---|---|
| One tool's headline numbers | posthog:query-mcp-tool-stats |
| One tool's day-by-day trend | posthog:query-mcp-tool-daily-stats |
| One tool's top errors | posthog:query-mcp-tool-failures |
| One tool's top callers | posthog:query-mcp-tool-top-users |
| Tools called before/after one tool | posthog:query-mcp-tool-neighbors |
| One tool's recent agent intents | posthog:query-mcp-tool-sample-intents |
| One tool's registered descriptions | posthog:query-mcp-tool-descriptions |
| Usage split by client harness | posthog:query-mcp-harness-breakdown |
| List sessions | posthog:mcp-analytics-sessions-list |
| One session's tool calls | posthog:mcp-analytics-sessions-tool-calls |
Not covered — use posthog:execute-sql:
- Cross-tool rankings (the tool-quality matrix — "which tool errors most?")
- Errored-session filtering (the session list has no error filter or error count)
- Effective tool names inside a session (
posthog:mcp-analytics-sessions-tool-callsreturns the raw$mcp_tool_name, not the inner tool of a wrapper call) - Any custom breakdown
posthog:execute-sql is also the fallback when the tools in the table above aren't
in your tool list. Query $mcp_tool_call directly; the schema and recipes are in
models-mcp.md.
Related skills
exploring-mcp-tool-quality— rank tools by error rate / latency / reach, then drill inexploring-mcp-sessions— a single agent run and its tool sequenceexploring-mcp-intent-clusters— agent goals grouped by semantic similarity