Data & privacy
Session Exporter reads your local session files directly. Nothing leaves your machine — there is no account, no telemetry, and no network call except serving the UI to your own browser on 127.0.0.1.
Where the data comes from
| Source | Location | Token basis |
|---|---|---|
| Claude Code | ~/.claude/projects/<folder>/<uuid>.jsonl |
recorded — summed usage, including cache create/read |
| Codex | ~/.codex/sessions/**/rollout-*.jsonl (and archived_sessions/) |
recorded — the final token_count event (includes cached input & reasoning) |
| Cursor | the global SQLite DB …/Cursor/User/globalStorage/state.vscdb |
context-snapshot — see below |
| Antigravity | ~/.gemini/antigravity{,-cli}/conversations/*.db |
recorded — gen_metadata usage (input, cache read, output, reasoning) |
| Pi Agent | ~/.pi/agent/sessions/**/*.jsonl |
recorded — per-turn usage (input, output, cacheRead, cacheWrite) |
| Kimi Code | $KIMI_CODE_HOME/sessions/**/ (default ~/.kimi-code) |
recorded — usage.record (uncached input, output, cache read/create) |
Claude Code
Each project folder holds one JSONL file per session. Token usage is summed from every assistant turn's usage object, and the 5-minute / 1-hour cache-write split is read out for accurate cost. The project path comes from the cwd recorded in the log.
Codex
Rollout files can reach hundreds of megabytes. Session Exporter streams them — counting messages at the byte level and JSON-parsing only the head (session metadata, first user message, model) and tail (final token count) — so a 650 MB session lists in well under a second. Message counts on very large sessions are a fast byte-level estimate (±a couple of system turns); every other number is exact.
Cursor
Cursor stores everything in one large SQLite database. Session Exporter opens it strictly read-only (?mode=ro, falling back to ?immutable=1) and only ever does indexed primary-key lookups — it never writes to, locks, or otherwise touches your live Cursor database, even while Cursor is running.
Cursor does not record cumulative token spend or cache activity — only the size of a session's final context window. Session Exporter surfaces that number honestly as context-snapshot (marked with ~) and does not compute a cost for it, rather than inventing one.
Antigravity
Antigravity (IDE and CLI) stores one SQLite database per conversation under ~/.gemini/antigravity/conversations/ and ~/.gemini/antigravity-cli/conversations/. Session Exporter opens each DB read-only (preferring ?immutable=1 so WAL sidecars never block the scan) and decodes the protobuf blobs in gen_metadata / steps / trajectory_metadata_blob with a tiny wire-format reader — no third-party protobuf dependency.
Token totals come from recorded generation metadata (uncached input, cache reads, output text, and reasoning). Workspace paths come from the trajectory metadata. Legacy encrypted .pb conversation files are skipped; only the SQLite .db format is supported.
Pi Agent
Pi Agent stores one JSONL file per session under ~/.pi/agent/sessions/<encoded-cwd>/. Each file opens with a type:"session" header (id, cwd, timestamp), then a stream of message events. Assistant turns carry recorded usage with input / output / cacheRead / cacheWrite. Session Exporter sums those fields and renders text, thinking, and toolCall blocks into the transcript.
Kimi Code
Kimi Code stores session metadata in state.json and the complete ordered event stream in agents/<agent>/wire.jsonl under $KIMI_CODE_HOME/sessions/ (default ~/.kimi-code). Session Exporter reads the main Agent and every sub-Agent log, merges their transcript events by timestamp, and sums only canonical usage.record events so generation usage is never double-counted.
The normalized fields inputOther, output, inputCacheRead, and inputCacheCreation provide exact token and cache accounting. The configured KIMI_CODE_HOME is respected. Known host instruction wrappers are removed from user-facing titles and transcripts; ordinary Kimi CLI prompts are left unchanged.
The cache
The scan result is written to .cache/index.json next to the app, keyed by a fingerprint of your session files (paths + sizes + modification times). If nothing changed, relaunches load from cache instantly. If files changed, or you click Refresh, the index is rebuilt. Deleting .cache/ just forces a fresh scan — it's safe.
Cost is applied to the cached index on every load, so editing pricing.json takes effect on the next launch (or immediately via ↻ Reload prices) without re-parsing your data.
Adding another source
Each parser is a single file in parsers/ implementing a small contract — list_sessions() and load_messages(). Adding another tool is one new file; nothing else in the app needs to know about it.