Conversations, events, users and analytics

Inspect transcripts and event timelines, tail the live event stream, look up end users, and read usage and cost.

Conversations

Filter by agent and status (Open, Handoff, Closed) in the selected environment. Columns: user, agent, version, title, messages, status, cost, last message, and (beta) the tags Jev assigns after each assistant turn: topic (question, purchase, support issue, navigation, feedback, onboarding, other), sentiment and whether the conversation looks resolved. Click a row for the inspector:

• Header facts: user, agent and version, status, usage (tokens in and out, cost on your key), activity (messages, tool calls, UI blocks), the user’s context, and the running summary of the messages folded out of the model’s window once the conversation is long enough.
• The transcript with schematic renderings of every surface.
• The event timeline: every kletso.events/v1 envelope with relative timing. Click a row for the JSON (tokens, secrets and authorization values are redacted). judge.turn rows (beta) show what Jev decided before the model ran: the intent, the option it read from the UI on screen, probabilities, latency and tokens.
• Workflow runs linked to the conversation.
• Close conversation.

Events refresh every few seconds while the inspector is open, so you can watch a live session.

Events

A live tail of platform events in the environment, filterable by family (message.*, tool.*, ui.*, app.*, trigger.*, error), with Pause and Resume. Each row links to its conversation; clicking shows the raw envelope. The full type list is in the events reference.

Users

End users the SDK has identified or seen anonymously: type (anonymous or identified by host JWT), plan trait, push platform, conversation count, first and last seen. The drawer shows traits, the runtime context snapshot and Memory: the durable facts the agent remembers about this user across conversations (how it works). Remove a fact, add one, or forget everything; changes apply to the next turn and are audited. The heading also shows how many snippets of the user’s earlier conversations are indexed for recall; forget everything deletes those too.

Analytics

Range of 7, 14 or 30 days. Cards: conversations, messages, tool calls, notifications, error rate, latency p50/p95, tokens in/out, estimated cost on your key.

Cost per day. Today, month to date, the selected range and the average per day, a stacked bar per day split by agent, tables by agent and by model (with the share of cached input), a daily table and a CSV download for finance. The runtime writes one rollup row per agent, model and day on every completed turn (text and voice; voice shows as voice:<model>), and the figures are estimates at the provider’s list price on your key with cache reads counted at a tenth. Rows from before 2026-10-09 were backfilled from conversations and are attributed to the day each conversation started. Four charts (conversations and messages; tool calls, errors and notifications; latency; tokens and cost), each with a data table fallback. Numbers come from the daily usage rollups the runtime writes on every completed turn. A Jev judgments card (beta) shows judgment calls, tokens and their estimated cost at TypeSafe’s list price; this runs on Kletso’s key and is informational, not billed to you.

Cost estimates

message.completed carries usage and costMicros (micro-dollars) computed from a per-model price table. Unknown models use a default rate, so treat the figure as an estimate and reconcile with your provider’s bill.