Long sessions
3.1M3 sessions ended several times larger than the 128K this machine usually reaches. Every turn after that re-reads the whole window.
acme-web · 214 turns · 62.4K → 604.1Kbilling-api · 168 turns · 58.1K → 511.7K
Local · open source · zero dependencies
Get more out of every Claude Code session.
Every Claude Code session on one page — which one is waiting on you, what each is spending, and how much of your limits is left.
npx claude-code-session-trackerThen open the printed http://127.0.0.1:3099.
3stars
170downloads / month
v0.8.0latest
Node 20 or newer · macOS, Linux, and Windows
Preview
The page it opens
Limits on top, sessions below, and details for any row you open.
Claude Code
Sessions
LiveUpdated 16:49:12
Session limit 14:30 → 19:30
Resets 19:30 · 2h 41m left
58%
- Used
- 1.9M
- At this pace
- Limit 18:52
At the last hour's pace you'll hit the limit at 18:52, 38m before it resets. Anthropic's reading of this five-hour window, 2m ago.
Weekly limit 18 Aug → 25 Aug
Resets Mon 25 Aug, 03:00 · 2d 10h left
59%
- Used
- 31.4M
- At this pace
- 88% at reset
Every model. Anthropic's reading of this week, 2m ago.
Last 7 days
Standing context
1.6MSessions open holding 74.3K — the system prompt, tools, skills and memory files, paid before anything is asked.
38 sessions
Ranked by what each accounts for, in the tokens a limit counts. A finding appears only where the range runs heavier than this machine's own middle.
Active 3
| Project | Session | Status | Model | Tokens (in/out/total) | Uptime |
|---|---|---|---|---|---|
| acme-web~/code/acme-web | Rework the checkout flow279ed6ae | busy | opus-5 | 182.4K / 12.9K / 195.3K | 1h 24m |
| billing-api~/code/billing-api | Paginate the invoices endpointb41c7d02 | waitinginput needed | sonnet-5 | 96.2K / 7.4K / 103.6K | 18m 42s |
| design-system~/code/design-system | Tidy the button variants5f2a91c4 | idle | sonnet-5 | 41.8K / 3.1K / 44.9K | 3h 06m |
Recent 5
| Project | Session | Model | Last active | Tokens (in/out/total) |
|---|---|---|---|---|
| acme-web~/code/acme-web | Fix the flaky checkout testit only fails on CI, never on my machine | opus-5 | 12m ago | 214.7K / 15.3K / 230.0K |
| mobile-app~/code/mobile-app | Retry the offline write queueflush the queued writes once the connection is back | sonnet-5 | 2h ago | 158.9K / 11.2K / 170.1K |
| docs-site~/code/docs-site | Add dark mode to the docsauto, light and dark, and no flash on first paint | sonnet-5 | 5h ago | 88.4K / 6.1K / 94.5K |
| billing-api~/code/billing-api | Rotate the webhook secretmove it to the secret store and re-sign the callbacks | haiku-4-5 | yesterday | 23.1K / 1.9K / 25.0K |
| acme-web~/code/acme-web | Windows CI: separators in the path testthe path assertions only hold on posix | opus-5 | 3d ago | 61.2K / 4.4K / 65.6K |
50 of 794 on disk
Read-only. Bound to loopback. Nothing leaves this machine.
/filter↑↓move↵openEscclose
Open any row
Tokens, context left, prompts, and a resume command to copy.
Static context shows what fills the window before you type — CLAUDE.md, skills, tools, MCP — the part you can actually shrink.
Every figure comes straight from the transcript. Nothing is guessed.
acme-web
Rework the checkout flow
busy
Tokens
- Input
- 182,411
- Output
- 12,907
- Total
- 195,318
- Cache read
- 4,182,660
- Cache write
- 311,204
Context window
- Static (system + tools)
- 41.2K
- Conversation
- 154.1K
- Free
- 804.7K
19% of 1M tokens · opus-5
Static context
- ~/code/acme-web/CLAUDE.md18,214 · 44%
- Skills (32)9,431 · 23%
- Deferred tools (86)5,102 · 12%
- Agents (11)3,608 · 9%
- MCP instructions2,417 · 6%
- Everything else2,435 · 6%
41,207 tokens before the first turn ran, estimated from the text each block put in the window.
What this session cost
Its 38 turns read 4.2M back out of cache between them. Those reads bill at a fraction of the rest, so they sit outside the totals above.
- The window an average turn held
- 110.1K
- Of it, static — read back every turn
- 41.2K
- Subagents, billed here but not shown
- 2
Session
- Model
- opus-5
- Started
- 14:12
- Last active
- just now
- Elapsed
- 1h 24m
- Working time
- 38m 09s
Limits
How much of each window is left
Two windows, five hours and seven days. Each card shows how full it is and when it runs out at your current pace.
The same percentage /usage shows
Anthropic's own reading of each limit, the same number /usage shows.
Used is what you were billed for
Input, output and new cache tokens, across every project and subagent. Cache reads are shown apart.
At this pace, it runs out at…
Based on your recent pace, the card says when the window runs out — or where it will stand at reset.
Two ticks on the bar
Red is time elapsed, blue is where you'll land. Fill past the red tick means you're spending too fast, and the bar turns red.
Offline or without a fresh reading, the bar falls back to your heaviest past window — and says so.
Alerts
Told before you hit a limit
A desktop notification before a window runs out. Off by default.
http://127.0.0.1:3099/settings
Claude Code
Settings
SessionsHistorySettingsNotifications
A desktop notification before a window runs past its limit.
Session limitThe five-hour window.
At most one every1 hourWeekly limitThe seven-day window.
At most one every4 hoursSession limit
now127.0.0.1:3099
At the last hour's pace you'll hit the limit at 18:52, 38m before it resets.
Hear about a limit before you reach it
A desktop notification when a window is on course to run past its limit — even with this tab behind your editor.
Not nowOpen notifications
Off until you turn it on
Nothing asks for permission until you flip a switch.
Once, then quiet
One per interval: an hour for the session, four for the week.
Offered once
Suggested once when a limit first appears.
A week with no reset from Claude stays silent.
History
Where the tokens actually went
A second page with your last month of tokens by day, hour, project and model.
http://127.0.0.1:3099/history?range=30d
Claude Code
History
SessionsHistorySettingsRangeLast 30 daysRefresh
25 Jul – 23 Aug · 30 days
- Billed
- 41,812,904
- Cache reads
- 2,012,517,338
- Turns
- 10,211
- Projects
- 5
- Models
- 3
- Busiest day
- 12 Aug · 2.9M
Spend per day 26 active
Billed tokens per local day. A mark under a bar is a turn Claude refused.
Hour of day
Every half hour of the range, over one week. Darker is heavier.
Projects 5
Ranked by billed tokens. Pick one to narrow everything above.
| Project | Turns | Billed | Cache reads | Share |
|---|---|---|---|---|
| acme-web~/code/acme-web | 4,182 | 18.4M | 902.6M | 44% |
| billing-api~/code/billing-api | 2,140 | 9.1M | 431.7M | 22% |
| mobile-app~/code/mobile-app | 1,884 | 6.7M | 318.2M | 16% |
| design-system~/code/design-system | 1,102 | 4.3M | 205.9M | 10% |
| docs-site~/code/docs-site | 903 | 3.3M | 154.1M | 8% |
Models 3
Every model that answered in this range.
| Model | Turns | Billed | Cache reads | Share |
|---|---|---|---|---|
| opus-5 | 6,204 | 27.6M | 1.3B | 66% |
| sonnet-5 | 3,411 | 12.1M | 620.4M | 29% |
| haiku-4-5 | 596 | 2.1M | 92.1M | 5% |
Read at 21:14
Spend per day, quiet days included
A bar per day. A mark under a bar is a day Claude refused a turn.
Hour of day, a month folded onto one week
A month folded onto one week — when your window actually opens.
Every project and model, ranked by what it billed
Ranked by billed tokens. Pick a project to narrow everything to it.
The range and the project live in the URL
Bookmarkable: range and project live in the URL. Capped at 90 days.
Reuses the limit cards' read, so it costs almost nothing. Nothing polls — click Refresh.
What you get
Everything you need to keep sessions moving
Limits, live status, token usage and history, on one page.
Quickstart
Run it with any package manager
No install needed. Or use Homebrew to keep it on your PATH.
npx claude-code-session-trackerRun it
No install step. It starts a local server and opens your browser.
Open the printed address
Served at http://127.0.0.1:3099.
Watch your sessions
Refreshes every 2 seconds. Click a row for details.
Start at login
NewOpen your dashboard every time you log in
One command adds it to your login items. Another takes it out.
brew install meyusufdemirci/tap/claude-code-session-trackerclaude-code-session-tracker autostart onAdded as a Login Item, listed in System Settings → General → Login Items.
To turn it off:
claude-code-session-tracker autostart offInstall it once
A login item needs a copy that stays on disk, so npx won't do here.
Turn it on
One command, no admin rights. It starts the tracker right away, too.
Log in
The dashboard opens at http://127.0.0.1:3099 every time.
Command line
Options
Every flag, command and keyboard shortcut.
| Flag or command | Description | Default |
|---|---|---|
| -p, --port <number> | Port to listen on, stepping forward up to 20 times if taken | 3099 |
| --host <address> | Address to bind. Anything but loopback drops the guard, and the CLI says so | 127.0.0.1 |
| --no-open | Do not open a browser | — |
| --json | Print the session list as JSON and exit | — |
| -n, --limit <number> | How many sessions to list. Running sessions always show | 50 |
| --claude-dir <path> | Override the Claude data directory | ~/.claude |
| --offline | Never ask Anthropic's server for the usage limits; the cards fall back to Claude Code's cached readout or an estimate | — |
| autostart on | Start at login and open the page, on macOS and Windows. Starts it now, too | — |
| autostart off | Stop starting at login | — |
| autostart status | Say whether it starts at login | — |
| -h, --help | Show usage | — |
| -v, --version | Show the version | — |
On the page
- /Jump to the filter
- ↑↓Move between sessions, across both tables
- HomeEndFirst and last session
- ↵Open the selected session
- EscClose the panel, or clear the filter
Scripting
An HTTP API, and JSON on stdout
Anything the page does, you can do with curl.
| Method | Route |
|---|---|
| GET | /api/sessions?limit=NThe session list. `limit` matches --limit, and running sessions are always included. |
| GET | /api/sessions?since=&until=The same list, narrowed by last-written time. Epoch ms; `since` inclusive, `until` exclusive. |
| GET | /api/sessions?sort=`recent` (default), `tokens-desc`, or `tokens-asc`. Ranks across the whole window, not just the page. |
| GET | /api/sessions/:idOne session with counts, tokens, models, activeMs, awaySummary, and notes. |
| GET | /api/limitsBoth limits, as `session` (five hours) and `weekly` (seven days). Each carries the window in progress, Anthropic's percentage as `reported`, the heaviest closed window, and `lastLimited`. |
| GET | /api/usage/history?since=&until=&project=A sparse half-hour series, plus every project and model in the range ranked by billed tokens. Defaults to 30 days; spans over 90 are narrowed, and `range` in the reply is what was read. |
| GET | /api/usage/adviceThe findings behind the `Where it goes` panel as measurements, not sentences. Same params as the history route; defaults to 7 days. |
| GET | /api/healthStatus, version, the Node it runs on, the resolved Claude directory, and per-source status. |
| POST | /api/sessions/:id/revealShows that transcript in your file manager. Requires a loopback Origin. |
--json prints the same payload the page uses, then exits.
npx claude-code-session-tracker --json | jq '.sessions[] | .project.name'{
"sessions": [
{
"id": "279ed6ae-49fd-4234-a74e-145f5535341c",
"source": "claude-code",
"status": "busy", // busy · waiting · idle · ended
"project": { "name": "…", "path": "…", "gitBranch": "main" },
"title": "Disable dependabot", // Claude's own title, when it wrote one
"firstPrompt": "…",
"lastPrompt": "…",
"model": "claude-sonnet-5",
"version": "2.1.235",
"startedAt": 1787142489923,
"lastActiveAt": 1787142700231,
"transcriptPath": "…/279ed6ae….jsonl",
"sizeBytes": 136133,
"live": { "pid": 4129, "kind": "interactive", "entrypoint": "cli" }
}
],
"total": 794,
"generatedAt": 1787142701002
}Privacy
Your transcripts never leave this machine
Read-only, loopback-only, and no telemetry.
Never writes to your Claude directory
It only reads what Claude Code already wrote.
Loopback only
Binds to 127.0.0.1 and rejects other hosts, so no website can reach it.
One outbound call, and --offline
Only Claude Code's usage endpoint. No telemetry, no analytics.
No dependencies, no install scripts
Nothing third-party runs on your machine.
What it reads
| Path | Used for |
|---|---|
| ~/.claude/sessions/<pid>.json | Running sessions and their live status |
| <session cwd>/.git/HEAD | The branch a running session is on |
| ~/.claude/projects/**/*.jsonl | Session history — titles, prompts, models, branch |
| ~/.claude/projects/*/*/subagents/agent-*.jsonl | Subagent turns, for the limit windows they bill to |
| ~/.claude.json | The usage readout Claude Code caches — how full each limit is, and when it resets |
| Keychain · ~/.claude/.credentials.json | The token Claude Code is signed in with, to ask for the same readout /usage shows. Skipped with --offline |
FAQ
Questions, and the honest answers
One command. No more guessing which session needs you.
Node 20 or newer · macOS, Linux, and Windows. MIT licensed.
npx claude-code-session-tracker