Cursor

Confabulous has first-class support for Cursor agent sessions. Cursor records a slimmer transcript than other providers, so a few cards behave differently — this page describes exactly what Confabulous reads and what it cannot.

What gets parsed

• Conversation history: user and assistant text from the main agent thread.
• Tool calls — the tool name plus a one-line summary of its input.
• File reads and edits, inferred from those tool calls.

Cursor’s agent transcript records tool inputs only, so there are no tool outputs to display. Confabulous shows each tool call and its arguments, but not the result the tool returned.

Cursor hides some scaffolding in its own UI and writes a bare [REDACTED] placeholder into the transcript instead; Confabulous strips that on-disk placeholder so transcripts read cleanly. This is Cursor’s own scaffolding, not a redacted secret — it is unrelated to the [REDACTED:TYPE] markers the confab CLI writes when it scrubs secrets on other providers.

Analytics cards

Each Cursor session produces the same provider-agnostic cards as every other provider:

• Tools — which tools were called and how often, using Cursor’s own tool names (Read, Grep, Glob, SemanticSearch, StrReplace, Write, Shell, Delete, WebSearch, AskQuestion, Task).
• Conversation — turn structure and message counts. Cursor’s transcript has no per-message timestamps, so per-turn timing and assistant utilization can’t be computed; the card shows only the metrics it can derive (user prompts) plus a short note explaining why the timing rows are absent.
• Code activity — files touched and language breakdown, derived from Cursor’s file tool calls.
• Agents & skills — subagent (Task) invocations, plus activity aggregated from subagent threads; see Subagents below.
• Session — high-level session metadata.

There is no Tokens or Cost card for Cursor — see Tokens and cost.

Tokens and cost

Per-session token counts and cost are not available for Cursor sessions. This is a limitation of the data Cursor stores locally, not a gap Confabulous can backfill:

• Cursor meters tokens server-side. The local agent transcript Confabulous ingests has no usable token counts — Cursor staff have confirmed the local tokenCount value is best-effort and unreliable.
• Cursor billing is subscription and usage-pool based, so there is no per-message dollar figure to attribute to a session.

Instead of hiding the Tokens card or showing misleading zeros, Confabulous shows an info callout on the Session Summary tab and “Not available” placeholders (with tooltips) on token and cost rows. The session list cost column shows an em dash rather than $0.00. On Trends, when Cursor sessions are in the selected window, the Tokens & Cost card shows the same unavailable state for Cursor’s section and an ⓘ caveat on the card title.

Real Cursor cost analytics, sourced from Cursor’s Dashboard / Admin API, are planned as future work — they are not part of this release.

Model

Cursor’s agent transcript does not record a per-message model identifier, so Confabulous does not show a model for Cursor sessions yet. Model attribution is planned for a future release.

Subagents

Cursor spawns subagents into separate transcript files. Confabulous uploads and parses those subagent files and aggregates their activity into the parent session: tool calls, code activity, agent invocations, and search text all include the subagents’ work.

Some cards stay main-thread only by design, so they reflect what you actually saw in the session:

• Conversation turn counts, message counts, session duration, and model reflect the main thread — subagents run within the main session, so they don’t widen these.
• The transcript pane shows the main thread only; subagent threads contribute to the cards and to search, but are not rendered as separate transcript rows.

Because subagent text feeds the search index, searching for a phrase that appears only inside a subagent will still match the session.

Deep links

Cursor’s transcript has no stable per-message or per-tool identifiers, so deep-linking directly to a specific Cursor message is not supported. Within the app the transcript pane renders normally; it just can’t be anchored to an individual message the way other providers can.

Other supported providers

Confabulous treats every provider as a first-class citizen. Claude Code, Codex, and OpenCode are also supported today. New providers slot into the same sync, storage, and analytics pipeline.