Local web dashboard that tracks context pressure across your AI coding agents — Cursor, Claude Code, and Codex — so you can review how context use has played out over time and adjust your strategy in future conversations.
- Cursor — reads
state.vscdb(cursorDiskKV → messageRequestContext:*), shows each composer's snapshot size broken down into buckets (webReferences,diffsSinceLastApply,projectLayouts,knowledgeItems, etc.) with cause tags (user-attached vs auto-injected vs mixed) - Claude Code — parses
~/.claude/projects/**/*.jsonltranscripts, surfaces token usage per conversation and per turn - Codex — reads the Codex SQLite DB (best-effort)
All three are normalized into a unified model so you can compare projects and conversations side-by-side, sorted by heaviest context first.
- Node 20+
- pnpm
pnpm install
pnpm devOpen http://localhost:3000.
JSON export (same data as the UI): http://localhost:3000/api/context
Query parameters for /api/context:
| Parameter | Default | Description |
|---|---|---|
db= |
macOS globalStorage path | Path to Cursor state.vscdb |
redact=1 |
off | Redact path-like strings before parsing |
cursor=0 |
on | Disable Cursor ingestion |
claude=0 |
on | Disable Claude Code ingestion |
codex=0 |
on | Disable Codex ingestion |
Save a snapshot to disk:
curl -s 'http://localhost:3000/api/context?redact=1' -o report.jsonProduction
pnpm build
pnpm start- Cursor's schema is undocumented and may change between releases.
- Cursor rows measure serialized snapshot size (bytes), not token counts. Claude Code and Codex use token-ish fields from transcripts or SQLite — not directly comparable.
- Read Cursor's DB while Cursor is idle so the WAL is checkpointed, or copy
state.vscdb*and passdb=to point at the copy.
See CONTRIBUTING.md.
MIT