kubit-inspect

Find and investigate Kubit entities — users, sessions, traces, and events. Use it to debug failures, drill into specific records, or navigate between related entities.

What it does

kubit-inspect is your tool for specific lookups and investigations. You point it at an entity (or a filter), and it returns a summary plus the matching records.

Kubit entities are hierarchical: user → sessions → traces → events

You can start at any level and navigate down. Inspect a user to see their sessions; inspect a session to see its traces; inspect a trace to see its events.

Prerequisites

A connected workspace. Run kubit-connect first if you haven't already.

Usage

❯ /kubit-inspect [<natural language query>]

You can write the query however feels natural — entity type, filters, and time ranges are all parsed for you. No need to memorize flag syntax.

Context: I created a funnel, here is its output.

kubit-inspect user IDs in the funnel

Examples

Look up a specific user

❯ /kubit-inspect user alex@acme.com

Returns a summary for that user: total cost, latency, token usage, top errors, session count, and most common intent. From there you can drill into their sessions or traces.

Navigate between entities

❯ /kubit-inspect sessions for alex@acme.com

❯ /kubit-inspect traces in session sess_abc123

❯ /kubit-inspect events fired in trace 1cd9d647

Follow-up queries like these work naturally — kubit-inspect carries forward the entity you were just looking at.

Filter traces

❯ /kubit-inspect failed traces with intent Checkout since yesterday

Returns a pattern summary across the matching traces (avg cost, avg latency, common error, common model) followed by a list of individual traces. The total match count is always shown alongside the number displayed.

Ask follow-up questions

❯ /kubit-inspect which of those traces had latency over 5s?

Follow-up questions that narrow or re-cut the same dataset reuse the cached results — they're faster and don't re-hit the MCP.

Example output for breaking down by failed traces:

Drill into a report segment

If you've just run a kubit-report (or pasted a report URL into the chat), kubit-inspect can pick up that context and zoom into the segment you're asking about.

❯ /kubit-inspect users who dropped off at payment in that funnel

Output behavior

• Single entity lookups return a conversational summary of the key fields — cost, latency, tokens, status, errors — with context for what they mean.

• Multi-result queries lead with the pattern (what's broken, expensive, or slow) and follow with a compact list of individual records.

• Large datasets are analyzed against the full export rather than a sample, so the numbers reflect everything that matched, not just the first ~100 rows.

Suggested next steps

After an inspection, you'll typically see a follow-up suggestion:

These are suggestions only — nothing runs automatically.

Notes

• If a query returns no results, try broadening the time range or double-checking filter values (intent names, user IDs, etc. are case-sensitive).

• Some query types don't support full-dataset analysis. When that's the case, you'll see a note that the summary is based on a limited sample.

• If a query is ambiguous about which entity type you mean (users, sessions, traces, or events), you'll be asked to pick rather than have one guessed for you.