Fast, local Claude Code log viewer. CLI, TUI, and static web — Rust binary, <1.3 MB.
ccaudit daily # token usage + cost, today and lately
ccaudit tui # interactive terminal browser
ccaudit web # generate + serve the web dashboard
ccaudit statusline # one-line summary for your shell prompt
One Rust binary. CLI, TUI, and web for your ~/.claude logs. Mmap'd cache, ~10 ms warm start.
Inspired by:
| ccaudit | ccusage | claude-code-log | claude-session-dashboard | |
|---|---|---|---|---|
| Runtime | Rust binary (~1.3 MB) | Node.js | Python | Node.js + browser |
| CLI reports | yes (daily/monthly/…) | reference impl | — | — |
| TUI browser | yes (ccaudit tui) |
— | — | — |
| Web dashboard | yes (ccaudit web) |
— | HTML export | yes (local server) |
| Session detail viewer | yes | — | yes | yes |
| Install | binary / npm / cargo | npm | pip install + py |
npx / npm install -g |
claude-session-dashboard has one thing ccaudit doesn't: an agent-delegation Gantt chart for the sub-agent dispatch tree per session. Worth it if you run agentic workflows and want to see the delegation order.
Same dataset, same workload (daily token report). Single-shot wall time on Apple Silicon, against ~/.claude/projects/ = 3.0 GB / 822 JSONL files.
| uncached | cold cache | warm cache | |
|---|---|---|---|
| ccaudit | 0.25 s | 0.07 s | 0.08 s |
ccusage (bunx) |
9.5 s | 8.9 s | 9.5 s |
claude-code-log (uvx) |
83.0 s | 52.2 s | 36.9 s |
- uncached — app-level cache wiped, forcing a full re-parse from JSONL
- cold cache — app cache present, OS page cache evicted (
sudo purge) - warm cache — re-run immediately afterwards; mmap + page cache hot
Same input (~/.claude/projects/), same logical question ("daily token totals"). Here's what each tool does for each invocation:
ccaudit daily — warm cache (~70 ms)
─────────────────────────────────────────────
mmap usage.db (3 MB binary, already bucketed by day × model)
│
▼
sum PreAgg cells
│
▼
print table
ccaudit daily — uncached (~250 ms, pays the cost ONCE per file lifetime)
─────────────────────────────────────────────
walk ~/.claude/projects → 822 files
│
▼
rayon-parallel parse JSONL (1.2 M lines, typed serde)
│
▼
bucket by (day × model)
│
▼
┌────────────────────────────┐
│ persist binary cache.db │ ◄── pay once,
└────────┬───────────────────┘ all future runs are 70 ms
▼
mmap + sum + print
ccusage daily (~9 s, EVERY SINGLE RUN)
─────────────────────────────────────────────
┌────────────────────────────────────────────┐
│ ⏬ HTTP GET litellm/prices.json (~1 MB) │ ◄── network call,
└────────┬───────────────────────────────────┘ every run
▼
parse pricing JSON (~1500 models)
│
▼
walk ~/.claude/projects → 822 files
│
▼
┌────────────────────────────────────────┐
│ for each of 1.2 M lines: │
│ ├── read line from disk │
│ ├── JSON.parse(...) │ ◄── re-parse every run
│ ├── lookup model in pricing map │
│ ├── tokens × cents │
│ └── push into per-day bucket │
└────────┬───────────────────────────────┘
▼
print table
claude-code-log (~37 s warm / ~83 s uncached)
─────────────────────────────────────────────
load Python pickle cache (if present)
│
▼
walk ~/.claude/projects → 822 files
│
▼
┌────────────────────────────────────────────┐
│ for each .jsonl: │
│ ├── parse JSONL │
│ ├── build parentUuid → child DAG │
│ ├── ⚠ orphan-node "promote to root" │ ◄── hundreds
│ ├── ⚠ multi-root fallback │ of these
│ └── ⚠ DAG incomplete → timestamp resort │
└────────┬───────────────────────────────────┘
▼
┌────────────────────────────────────────────┐
│ for each session: │
│ └── jinja2 template → write .html file │ ◄── thousands of
└────────┬───────────────────────────────────┘ output files
▼
write combined_transcripts.html
│
▼
update pickle cache
The redundancy in one row:
| once per run | ccaudit (warm) | ccusage | claude-code-log |
|---|---|---|---|
| stat 822 JSONL files | — | ✓ | ✓ |
| parse 1.2 M JSONL lines | — | ✓ | ✓ |
| build per-message parent→child DAG | — | — | ✓ |
| render thousands of HTML files | — | — | ✓ |
| HTTP-fetch LiteLLM prices | — | ✓ | — |
| look up model prices, multiply by tokens | — | ✓ | — |
| sum already-bucketed numbers + print | ✓ | ✓ | — |
ccaudit does the top six rows once, when it first sees a file, persists the result to a 3 MB binary cache, and every subsequent run is just the bottom row.
ccaudit keeps two purpose-built caches under ~/.claude/ccaudit-cache/. CLI reports, the TUI, and the web view each consume the one shaped for what they actually need:
~/.claude/ccaudit-cache/
├── claude-code.db ◄── aggregation cache (CLI reports)
│ bucketed by (day × model × project),
│ mmap'd, zero deserialization on read
│
├── codex.db ◄── same shape, per --source provider
│
├── <hash>.meta ◄── per-session cache (TUI + web)
└── <hash>.bin one pair per JSONL file —
full parsed Session struct, every
message, every tool call, bincoded
(validated by mtime + size + version)
CLI (daily / monthly / session / blocks / statusline) reads only <source>.db — no per-message data needed for token totals.
TUI (ccaudit tui) reads only the per-session .bin files. On startup, load_all_projects() walks every JSONL; for each file it stat's mtime + size, hits the matching .bin if those match the meta, and bincode-deserializes straight into a Session. A miss triggers a reparse + cache rewrite for next time. Once everything is in memory, navigation, search, dashboard, scope filters — all in-memory, zero IO. Pressing c exec's claude -r <id> against the current session.
Web (ccaudit web) reuses the same per-session cache for the parse step, then materializes the browser-side cache as static files under ~/.claude/ccaudit-web/:
~/.claude/ccaudit-web/
├── index.html
├── app.js / style.css / util.js ◄── single-file bundle (no build step)
├── index.json ◄── one fetch:
│ project + session metadata,
│ hourly histograms, tool counts,
│ + a daily rollup pulled straight
│ from <source>.db so the heatmap
│ can't drift from the CLI
├── search.json ◄── word → session posting list
└── s/
└── <pi>_<si>.json ◄── per-session message tree,
lazy-loaded by the browser
only when you open that session
The browser fetches index.json once and renders the dashboard / table from it. Opening a session is one HTTP GET for <pi>_<si>.json; repeat opens hit the browser's HTTP cache. The bundled HTTP server is a 10-line static file handler, no API.
End result: writing a daily report, scrolling the TUI, and clicking through 200 sessions in the web view all share the same parsed-once-on-disk substrate. The only thing that ever re-parses a JSONL file is a real change to that file (mtime or size moves).
Ephemeral (no install — runs the latest release once):
npx ccaudit # Node / npm
bunx ccaudit # Bun
uvx ccaudit # uv / AstralPermanent:
npm install -g ccaudit # npm / yarn / pnpm
cargo install ccaudit # cargo (from crates.io)
pipx install ccaudit # pipx (Python)
uv tool install ccaudit # uv tool
# from source
git clone https://github.com/electricapp/ccaudit && cd ccaudit
cargo install --path . --lockedAll five package managers resolve to the same pre-built platform
binary — there's no Node or Python runtime involvement at execution
time, the Python/JS shims just locate and exec the binary.
ccaudit # daily report (default)
ccaudit blocks --cost-limit 100 # 5-hour billing windows w/ progress bar
ccaudit session --breakdown # per-session per-model cost
ccaudit web --port 8080 # static site + local server
ccaudit web --no-serve --out ./site # generate the static site, exit (CI / scripts)
ccaudit tui # interactive browser
ccaudit --source codex daily # OpenAI Codex CLI logs (~/.codex/sessions)All subcommands accept --json for machine-readable output.
- Daily / monthly / session / blocks reports, plus a compact
statuslinefor shell prompts - TUI browser — keyboard-driven navigation, fuzzy search, message viewer, dashboard (
d), resume (c) - Web dashboard — tables with sortable columns, pie/histogram/heatmap charts, full message viewer, URL routing (
/p/{slug}/s/{uuid}) - Scope-aware: press
dfrom any view and the dashboard reflects just that project or session - Per-token-type cost breakdown on hover
- Accurate pricing — fetches latest rates from LiteLLM on demand (
ccaudit refresh-prices) - Carbon footer (
--carbon) — energy / CO₂ / tree-year estimate for the reported window - Deterministic filters —
--since YYYYMMDD,--until,--project,--timezone,--locale,--source - mmap'd cache — repeated runs on the same
~/.claude/projects/read from a memory-mapped schema, zero deserialization - Pluggable sources — a
Sourcetrait (seesrc/source/mod.rs) abstracts log discovery, parsing, model pricing, and model normalization. Today ships Claude Code (--source claude-code, default) and OpenAI Codex CLI (--source codex); the trait accepts adapters for OpenCode, π / Pi, MCP servers, or any other agent that writes JSONL-shaped session logs. - Single source of truth — CLI
daily, web sessions table, and web heatmap all read from onecache::per_session_totalspipeline. A test (tests/uniformity.rs) asserts they agree to the cent on every run.
cargo install ccaudit --features simd-json # ~30% faster JSONL parse, +300 KB binary
cargo install ccaudit --features locale # locale-aware date formatting (chrono unstable-locales)
cargo install ccaudit --features full # tui + web + locale (default omits locale)| key | action |
|---|---|
↑/↓ or k/j |
move selection |
Enter or → |
open |
← or Esc |
back |
/ |
search |
d |
toggle dashboard (scoped to current view) |
c |
claude -r <id> — resume this session |
o |
open the web view |
r |
reset all filters / scope |
q |
quit |
| key | action |
|---|---|
d |
toggle scoped dashboard |
m / t |
pie: by model / by tool |
h / p / y |
histogram: hour / project / day |
l |
histogram log/linear scale |
r |
reset all filters / sort / scope |
j/k |
row navigation |
ccaudit daily daily token usage + cost (default)
ccaudit monthly aggregate by month
ccaudit session aggregate by conversation session
ccaudit blocks 5-hour billing windows, with active detection
ccaudit statusline compact one-line summary (for terminal status bars)
ccaudit tui interactive TUI browser
ccaudit web generate static site + serve
ccaudit refresh-prices fetch latest model prices from LiteLLM
ccaudit help [SUB] show help for ccaudit or for a subcommand
Run ccaudit <SUBCOMMAND> --help for mode-specific flags.
cargo test --release --all-features # 41 tests across 5 suites
cargo run --release --example bench # synthetic-corpus bench
BENCH_SIZE=large BENCH_RUNS=10 \
cargo run --release --example bench # scaling stress
BENCH_SAVE=baseline.json cargo run --release --example bench
BENCH_COMPARE=baseline.json cargo run --release --example bench # diff vs baselineThe bench builds a deterministic JSONL fixture in a tempdir (no
dependency on ~/.claude/projects/) and times the macro paths
(parse cold, parse warm, cache rebuild, cache warm mmap,
cache::aggregate day & session, web::generate) plus a few inner
loops (parse_session, tokenize, Searcher::score, fnv1a).
MIT. See LICENSE.