Claudemeter

Monitor your Claude Code usage as you go.
Session, weekly and context usage live in your status bar, so you pace yourself instead of simply reacting. Context-rot and happy-hour awareness across every Claude plan.

• Token context usage
• Session limits
• Weekly limits
• Limit consumption and reset times
• Claude service status (working, partial outage, major outage)
• Happy hour indicator - lights up during Anthropic's off-peak window
• Context rot indication
• Claude session and login all local to device
• Open source: https://github.com/hyperi-io/claudemeter

---

Status Bar Default

Status Bar All Warnings and Happy Hour

Like Claudemeter?

If it saves you a surprise limit or a context-rot faceplant, a moment to star it on GitHub or leave a rating on the Marketplace would be genuinely appreciated. It helps other Claude Code users find it, and it tells us the work is worth carrying on. No pressure either way.

Claudemeter Browser Login No Longer Required

Claude Code changed - its CLI token can now reach the usage API, which it couldn't before. "It just works" now. Install and it works. So as of v2.5+ claudemeter reads that token directly and drops the browser entirely. This also fixes the Google SSO "browser not secure" block, since there's no automated login window any more (#49).

Context Window Detection

Claudemeter automatically detects your context window size - no manual configuration needed.

Anthropic's March 2026 GA rollout made 1M context the default for Max, Team, and Enterprise plans on current-generation models - Opus 4.6 and later (4.7, 4.8), Sonnet 4.6 and later, and Fable 5 (launched June 2026) - no [1m] suffix required. Pro and Free plans stay at 200K unless a user explicitly picks a [1m]-suffixed alias or tops up via pay-as-you-go.

Because Claude Code strips the [1m] suffix from model IDs before writing them to session logs, and because Claude Code's own s1mAccessCache can go stale, claudemeter can't rely on any single source. Instead it uses a priority chain and labels the result honestly in the tooltip:

1. User override - claudemeter.tokenLimit setting, if set (authoritative)
2. Explicit alias suffix - claudeCode.selectedModel: "opus[1m]" (authoritative)
3. JSONL suffix - a model ID with [Nm] in session logs (authoritative, rare in practice)
4. Live plan + model rule table - subscriptionType from the OAuth /profile response + model family from session logs matched against a data-driven rule table (e.g. Max + opus-4.6+ or fable-5 -> 1M)
5. Local plan + model rule table - same table, but plan comes from the token blob's subscriptionType (used before the first fetch or when web usage is off, e.g. tokenOnlyMode)
6. Claude Code's s1mAccessCache - used only as a last-resort corroborating signal, never as a negative
7. Observed usage snap-to-tier - if all authoritative signals fail but observed tokens exceed 200K, snap up to the next known tier (200K -> 1M -> 2M) and label the result as (inferred)
8. Standard fallback - 200K

The tooltip shows the source:

• Context: 1.0M - from an authoritative signal (user override, explicit alias, JSONL suffix)
• Context: 1.0M (configured) - from Claude Code's own eligibility cache
• Context: 1.0M (inferred) - from a rule-table match or observed-usage snap
• Context: 200K - standard fallback, no signal

The rule table is future-proof via numeric minVersion comparison - when Anthropic ships Opus 4.7 or 5.0 with the same defaults, the existing rules keep matching without a code change. To override auto-detection, set claudemeter.tokenLimit to a specific value.

Why the context-rot meter exists

The Tk gauge can turn light blue at ~400K tokens used, dark blue at ~650K, and yellow / red as Claude approaches auto-compact - before the existing yellow warning fires on a 200K-window account. These tiers exist because Claude's effective recall degrades long before the auto-compact trigger fires, and because the meaning of "X% used" changes dramatically with window size.

Update - Opus 4.8 - Jul 2026

The blue tiers track the current model. On Opus 4.8 they sit at ~400K (light) / ~650K (dark) on a 1M window - later than the 4.7-era 300K/500K, because 4.8 holds long context far better (it retains ~79% of its 256K retrieval score at 1M, vs 4.7's ~52%). There is still no third-party depth-binned 4.8 data between 256K and 1M, so the exact positions are a judgement call biased conservative for cache-heavy coding use. Full research record, anchor numbers, and the re-check log live in docs/context-rot.md - first researched 30 May 2026, re-checked 7 Jul 2026.

The apples-to-apples progression. GraphWalks BFS F1 is the only metric published for all three versions. Higher is better.

Model	BFS 256K	BFS 1M	keeps at 1M
Opus 4.6	61.1%	16.3%	~27%
Opus 4.7	76.9%	40.3%	~52%
Opus 4.8	85.9%	68.1%	~79%

What sparked this - Opus 4.7 - May 2026

Snapshot framing. Claudemeter shipped the rot meter in direct response to the Opus 4.7 multi-needle regression. The specific numbers below will shift as model implementations improve, degrade, or change - the meter itself stays useful because the phenomenon (long-context quality degradation) is durable. Treat this section as the spark and the historical reason, not a live spec.

From Anthropic's own Opus 4.7 model card, MRCR v2 8-needle retrieval at the time of this enhancement (May 2026). Different metric to the GraphWalks table above - the numbers do not compare. It stops at 4.7 because the 4.8 card dropped MRCR.

Context size	4.6 (MRCR)	4.7 (MRCR)
256K	91.9%	59.2%
1M	78.3%	32.2%

I wrote articles on how great the Opus 4.6 1m achievement was for 4.6 on 1m with so little rot... only to have on egg on my face the very next release with such a huge decline. I expect the costs were not acceptable for Anthropic.

Multi-needle recall - the realistic case when you're juggling several rules / files / constraints in one session - collapsed to roughly half of the 4.6 value at 256K, and to a third of it at 1M. Single-needle ("find this one thing") still scored ~89% at 1M on 4.7, but that's not how most real work uses the model.

Anthropic's own model card conceded Opus 4.6 with 64k extended-thinking dominates 4.7 on long-context multi-needle retrieval.

What this means in practice

The failure mode users describe most consistently:

"Claude does the right thing for the immediate ask but violates a rule from the original brief."

That happens before the existing yellow / red tiers fire. Light blue (~400K tokens used) and dark blue (~650K) are the new advisory tiers that flag this window - colour first, with a tooltip recommendation pointing at /compact on your terms.

How yellow / red defaults differ by account tier

The yellow and red Tk tiers are about one thing: how much runway you have before Claude compacts and drops its brain. Auto-compact is when Claude Code summarises the conversation to free up context; everything that doesn't survive the summary is lost from the model's working memory. You usually notice it as Claude suddenly forgetting a rule from earlier in the session.

Auto-compact is reserve-based, not a fixed percentage. It fires when only a small fixed number of tokens (~33K) remain in the window. The visible percentage at which it fires therefore depends on the window size:

Window	Auto-compact at
200K	~83%
1M	~96.7%

That's why Claudemeter uses profile-based thresholds keyed to your detected Claude tier. Yellow and red defaults are expressed as runway in absolute tokens (yellow = "~20K runway", red = "~5K runway"), so they mean the same thing across window sizes:

Window	Yellow fires at	Red fires at	Auto-compact at
200K	147K used (~74%)	162K used (~81%)	~167K (~83%)
1M	947K used (~95%)	962K used (~96%)	~967K (~97%)

Red consistently means "about 40K of runway left - Claude is about to drop its brain, finish your thought." Yellow means "60-80K of runway, still safe but heads-up." On a 200K window each percentage point is worth roughly 5x as much absolute runway as on 1M, which is why the percentage thresholds differ so much between the two - but the meaning (token runway to brain-loss) is identical.

The blue rot tiers (~400K / ~650K used) are about something different - quality degradation in long-context multi-step work - and apply on any window larger than 200K, because a big window is what makes the 400K / 650K frontier reachable in the first place. They are not auto-compact warnings, they are "context is getting long enough that Claude's recall is going to drift" warnings.

Built-in profiles

Each Claude account tier resolves to a built-in threshold profile:

Profile	Detected by	Typical window
pro	subscriptionType='pro'	200K
max-5x	rateLimitTier='default_claude_max_5x'	1M Opus auto
max-20x	rateLimitTier='default_claude_max_20x'	1M Opus auto
max-unknown	subscriptionType='max' alone	1M Opus auto
team-standard	orgType='Team'	1M Opus auto
enterprise	orgType='Enterprise'	500K
unknown	fallback	200K (assumed)

The profile sets the yellow / red runway. The blue rot tiers are NOT profile-gated - they fire whenever the detected context window is larger than 200K. So a 1M session shows rot regardless of which profile was detected, a 500K enterprise window shows light-blue rot, and a 200K session never does (400K is out of reach). Rot follows the window because tier detection isn't always reliable, but the window almost always is.

Customisation

• Force a profile: claudemeter.thresholds.tokens.profileOverride = "max-20x" etc.

• Override individual thresholds: edit claudemeter.thresholds.tokens.profiles deep-merge map. Example:

  "claudemeter.thresholds.tokens.profiles": {
    "max-20x": {
      "thresholds": {
        "warningRunwayTokens": 30000
      }
    }
  }

• Switch to monochrome (no threshold colouring): claudemeter.statusBar.colorMode = "basic".

• Custom hex per tier: claudemeter.colors.{rotLight, rotDeep, warning, error, happyHour}.

The rationale above stays in the README so users understand the why even if they choose to switch the meter off.

Happy Hour

Anthropic throttles Claude Code's 5-hour session window harder during their peak hours (Mon-Fri 05:00-23:00 America/Los_Angeles, per the 2026-03-26 announcement - see claude-code#41788 / #41930). Outside that window - weekday overnight and all weekend - the session token allowance burns at its expected rate.

Claudemeter lights up a dedicated status-bar panel when you're in that off-peak window, with a countdown to when peak kicks back in:

Claude ✨ 4h 17m  Se ●●○○○ ⌚ 1h 28m  Wk ●●○○○ ⌚ 4d 17h  Tk ●○○○○ 1m

Default icon is VS Code's monochrome $(sparkle) codicon - it inherits the status-bar text colour so it doesn't stand out. Shown as ✨ above because GitHub-rendered markdown can't display VS Code codicons directly. The real panel renders a small monochrome four-point star, not an emoji. You can swap it for a full-colour emoji (🍺, 🍹, ☕, etc.) via claudemeter.happyHour.icon.

The panel disappears entirely during peak - no empty slot. The countdown respects claudemeter.statusBar.timeFormat (default countdown -> 4h 17m, or 12hour / 24hour for a clock time).

Window is LA-local so the icon lines up with Anthropic's infrastructure peak, regardless of your own timezone. Override via claudemeter.happyHour.peakWindow if policy changes before a claudemeter release ships. Choose a different icon (or disable the panel entirely) via claudemeter.happyHour.icon / claudemeter.happyHour.enabled.

How It Works

Claudemeter reads the OAuth token Claude Code already stores and calls the same usage endpoints Claude Code's own /usage uses (api.anthropic.com/api/oauth/usage + /profile). Claude Code has evolved, so from 2.5 onwards no browser, no scraping, and no separate login are needed - if you use Claude Code, it just works.

The token lives in the shared Claude Code store: the macOS Keychain (Claude Code-credentials) or ~/.claude/.credentials.json elsewhere. Claudemeter reads it fresh each fetch and watches the store, so a token refresh is picked up automatically. It never refreshes or writes the token back, so it can't disturb your Claude Code login - and the usage always matches whichever account Claude Code is on.

Two independent streams feed the status bar. Web usage (session / weekly / opus / sonnet / credits) is account-global and shows in any window. The Tk context gauge is per-workspace - read live from the current project's Claude Code session - so an empty window (no project) shows the account gauges but no Tk. Multiple VS Code windows share a single web-usage fetch per claudemeter.usageRefreshSeconds interval, so N windows don't each hit the API and trip its rate limit.

Why not a browser? An earlier build tried this same OAuth approach first, but back then the Claude Code token wasn't authorised for the usage/account API - so the only option was a sessionKey cookie from a real browser login, which meant shipping browser automation and hitting Google's anti-automation SSO block. Anthropic has since granted the token that access, so browserless works now. The browser, the cookie, and playwright-core are gone. The extension ships with zero runtime dependencies.

Installation

Prerequisites

• VS Code 1.110.0 or higher
• Claude Code installed and logged in (the CLI or the official VS Code extension - they share one login). Claudemeter reads that login to show your usage.

First-Time Setup

If you already use Claude Code, there's nothing to set up - claudemeter reads your existing login and shows usage immediately.

If you're not logged into Claude Code, claudemeter prompts you:

1. Click Log in - claudemeter runs claude auth login in an integrated terminal.
2. Complete Anthropic's login in your real browser (SSO works - it's Anthropic's own login, not an automated window).
3. Claudemeter detects the new token and refreshes automatically.

If Claude Code isn't installed at all, claudemeter points you at the install docs instead.

Switching Claude Code accounts needs no action: the token in the shared store changes to the new account, and claudemeter's store watcher refetches automatically. Multiple VS Code windows running claudemeter at the same time are safe - the session-data file is locked and atomically merged so concurrent writers don't clobber each other.

Configuration

Open VS Code Settings and search for "Claudemeter" to configure:

claudemeter.fetchOnStartup

• Type: Boolean
• Default: true
• Description: Automatically fetch usage data when VS Code starts

claudemeter.usageRefreshSeconds

• Type: Number
• Default: 120
• Range: 30-3600 seconds
• Description: How often web usage (session / weekly / opus / sonnet / credits) is refreshed. The account-global usage is fetched once per interval and cached so all VS Code windows share one network call, rather than each hitting the Claude API and tripping its rate limit (429). Raise it if you still see rate-limit errors with many windows open.

claudemeter.localRefreshSeconds

• Type: Number
• Default: 10
• Range: 5-60 seconds
• Description: Local token refresh interval in seconds. Controls how often local Claude Code token data is polled from JSONL files. This is a low-overhead local operation (no web requests). Set higher to reduce CPU usage.

claudemeter.tokenLimit

• Type: Number
• Default: 0 (auto-detect)
• Range: 0-2000000
• Description: Context window token limit override. Set to 0 (default) to auto-detect from Claude Code's model selection. Set manually to force a specific limit.

claudemeter.sessionWindowMinutes

• Type: Number
• Default: 10
• Minimum: 1
• Description: The live window. The Tk gauge shows the largest session written within this many minutes, so concurrent sub-agent work reflects the heavy orchestrator, not a small sub-task.

claudemeter.sessionMaxAgeMinutes

• Type: Number
• Default: 30
• Minimum: 1
• Description: The hard deck - max last-modified age for a session file to count at all. The local Claude Code CLI drops an idle session after ~30min and loses its context, so past this a file is treated as dead.

How the two interoperate:

• Within sessionWindowMinutes (10 min): the largest live session wins.
• Older than that but within sessionMaxAgeMinutes (30 min): nothing is live, so the most-recent file's last value is shown (aged-out fallback).
• Older than sessionMaxAgeMinutes (30 min): all dead - the gauge blanks to Tk - and usage reads 0.

Keep sessionWindowMinutes <= sessionMaxAgeMinutes.

claudemeter.tokenOnlyMode

• Type: Boolean
• Default: false
• Description: Token-only mode - show only the local Tk context gauge, skip fetching web usage (session / weekly / opus / sonnet / credits) from the Claude API.

claudemeter.statusBar.displayMode

• Type: String
• Default: default
• Options: default, minimal, compact
• Description: Status bar display mode:
  • default: Full display with reset times (separate panels)
  • minimal: Percentages only (separate panels)
  • compact: All metrics in a single panel

claudemeter.statusBar.showSonnet

• Type: Boolean
• Default: false
• Description: Show Sonnet weekly usage in status bar (default/minimal modes only)

claudemeter.statusBar.showOpus

• Type: Boolean
• Default: false
• Description: Show Opus weekly usage in status bar (Max plans only, default/minimal modes)

claudemeter.statusBar.showCredits

• Type: Boolean
• Default: false
• Description: Show extra usage (spending cap) in status bar (default/minimal modes only)

claudemeter.statusBar.showServiceStatus

• Type: Boolean
• Default: true
• Description: Show Claude service status indicator. Displays a warning/error icon if Claude services are degraded or experiencing an outage.

claudemeter.happyHour.enabled

• Type: Boolean
• Default: true
• Description: Show the happy-hour status-bar panel during Anthropic's off-peak window. Set false to hide the panel entirely.

claudemeter.happyHour.icon

• Type: String
• Default: sparkle
• Options: sparkle, watch, zap, star, beer, cocktail, wine, champagne, martini, coffee, moon, sparkles, palm, party, custom
• Description: Glyph to render in the happy-hour panel. The first four are monochrome VS Code codicons (inherit the status-bar text colour, stay unobtrusive). The rest are full-colour emoji. Choose custom and set claudemeter.happyHour.customIcon to use any glyph.

claudemeter.happyHour.customIcon

• Type: String
• Default: ""
• Description: Custom glyph for the happy-hour panel. Only used when claudemeter.happyHour.icon is custom. Accepts any emoji or $(codicon-name) syntax.

claudemeter.happyHour.peakWindow

• Type: Object
• Default: { "days": [1,2,3,4,5], "start": "05:00", "end": "23:00", "tz": "America/Los_Angeles" }
• Description: Anthropic's peak-throttling window. Happy hour appears outside this window. days are 0 (Sunday) through 6 (Saturday). start / end are HH:MM 24-hour format in the given IANA tz. Override if Anthropic changes the policy before a claudemeter release ships. Malformed fields fall back to defaults individually (e.g. a bad tz still keeps your custom start / end).

claudemeter.statusBar.timeFormat

• Type: String
• Default: countdown
• Options: 12hour, 24hour, countdown
• Description: How to display reset times in the status bar:
  • 12hour: 12-hour format with AM/PM (e.g., 2:30 PM)
  • 24hour: 24-hour format (e.g., 14:30)
  • countdown: Countdown timer (e.g., 2h 15m)

claudemeter.statusBar.usageFormat

• Type: String
• Default: barCircle on macOS/Linux, barLight on Windows (platform-aware - Windows status-bar fonts often lack the Geometric Shapes glyphs barCircle uses, so the default falls back to Block Elements which render reliably). Any explicit user setting overrides the platform default.
• Options: percent, barLight, barSolid, barSquare, barCircle
• Description: How to display usage values in the status bar:
  • percent: Percentage (e.g., 60%)
  • barLight: Light blocks (e.g., ▓▓▓░░) - universal default for Windows
  • barSolid: Solid blocks (e.g., ███░░)
  • barSquare: Squares (e.g., ■■■□□)
  • barCircle: Circles (e.g., ●●●○○) - default on macOS/Linux

If your chosen style renders unevenly (mismatched cell widths), pick barLight or barSolid - Block Elements are the most font-portable option across all three platforms.

claudemeter.statusBar.tokensDisplay

• Type: String
• Default: limit (new in 2.3.3 - previously both, auto-migrated to extended)
• Options: bar, value, extended, limit, count
• Description: Controls the Tk (token usage) indicator:
  • bar: progress bar / percentage only - Tk ●●○○○
  • value: bar + current consumption - Tk ●●○○○ 518k
  • extended: bar + current/max - Tk ●●○○○ 518k/1m (same rendering as the pre-2.3.3 both default)
  • limit (default): bar + max, shown only when greater than the 200K standard context window - Tk ●●○○○ 1m on a 1M session, just the bar on a 200K session
  • count: token count only, no bar - Tk 518k/1m
• The max is suppressed when the context window limit is inferred rather than authoritative, so the display doesn't misrepresent an uncertain value.
• The existing claudemeter.statusBar.usageFormat setting still controls the bar/percent half.
• Migration: existing configs with tokensDisplay: both are auto-rewritten to extended on startup. No visible change for those users.

claudemeter.statusBar.alignment

• Type: String
• Default: right
• Options: left, right
• Description: Status bar alignment. Requires window reload to take effect.

claudemeter.statusBar.priority

• Type: Number
• Default: 100
• Range: 0-10000
• Description: Status bar priority (higher values position items closer to the centre). Requires window reload to take effect.

claudemeter.debug

• Type: Boolean
• Default: false
• Description: Enable debug logging to output channel (for troubleshooting)

claudemeter.debugLogFile

• Type: String
• Default: Auto-populated on first run (~/.config/claudemeter/debug.log or platform equivalent)
• Description: Path to debug log file. Supports ~ for home directory.

claudemeter.debugLogMaxSizeKB

• Type: Number
• Default: 256
• Range: 64-2048 KB
• Description: Maximum debug log file size in KB. Oldest entries are trimmed when exceeded.

claudemeter.thresholds.warning

• Type: Number
• Default: 80
• Range: 1-100
• Description: Usage percentage to show warning (yellow) indicator

claudemeter.thresholds.error

• Type: Number
• Default: 90
• Range: 1-100
• Description: Usage percentage to show error (red) indicator

claudemeter.thresholds.tokens.warning

• Type: Number
• Default: 65
• Range: 0-100
• Description: Token-gauge warning (yellow) threshold as a percentage. Default 65 tracks Claude Code's auto-compact trigger (~83% of the context window) minus ~20, giving roughly 20 percentage points of warning runway before auto-compact. 0 inherits claudemeter.thresholds.warning. Deprecated for profile-based accounts: the profile system expresses this as warningRunwayTokens (absolute tokens before auto-compact) - use claudemeter.thresholds.tokens.profiles to override.

claudemeter.thresholds.tokens.error

• Type: Number
• Default: 75
• Range: 0-100
• Description: Token-gauge error (red) threshold as a percentage. Default 75 gives roughly 10 percentage points of red runway before auto-compact fires (~83%). 0 inherits claudemeter.thresholds.error. Deprecated for profile-based accounts: use claudemeter.thresholds.tokens.profiles to override errorRunwayTokens instead.

claudemeter.statusBar.colorMode

• Type: String
• Default: color
• Options: color, basic
• Description: Status-bar decoration mode.
  • color (default): full palette - gauge tiers (rotLight, rotDeep, warning, error), Se/Wk warning icons, happy-hour green, and the colour-coded ● prefix in the tooltip.
  • basic: every gauge renders in default foreground - no icons, no tier tints, no tooltip prefix dots, no rot recommendation sub-lines. Tooltip section structure is preserved either way.

claudemeter.thresholds.tokens.profileOverride

• Type: String
• Default: "" (auto-detect)
• Description: Force a specific Tk threshold profile by name. Empty (default) auto-detects from your Claude account signals. Built-in profile names: pro, max-5x, max-20x, max-unknown, team-standard, enterprise, unknown. See Built-in profiles for what each profile enables.

claudemeter.thresholds.tokens.profiles

• Type: Object

• Default: {}

• Description: Per-profile threshold overrides (deep-merge semantics - leaf fields merge over the built-in profile, unset fields inherit from the built-in). Example overriding only the warning runway for max-20x:

  "claudemeter.thresholds.tokens.profiles": {
    "max-20x": {
      "thresholds": {
        "warningRunwayTokens": 30000
      }
    }
  }

claudemeter.colors.rotLight

• Type: String (hex colour or "")
• Default: "" (inherits theme colour claudemeter.rotLight)
• Description: Custom hex (e.g. #6ca0c4) overriding the light-blue rot tier colour. Theme-agnostic - applies to both the status bar and tooltip.

claudemeter.colors.rotDeep

• Type: String (hex colour or "")
• Default: "" (inherits theme colour claudemeter.rotDeep)
• Description: Custom hex overriding the dark-blue deep-rot tier colour. Theme-agnostic.

claudemeter.colors.warning

• Type: String (hex colour or "")
• Default: "" (inherits charts.yellow)
• Description: Custom hex overriding the warning (yellow) tier colour.

claudemeter.colors.error

• Type: String (hex colour or "")
• Default: "" (inherits claudemeter.outageRed)
• Description: Custom hex overriding the error (red) tier colour.

claudemeter.colors.happyHour

• Type: String (hex colour or "")
• Default: "" (inherits claudemeter.happyHourGreen)
• Description: Custom hex overriding the happy-hour panel colour.

Commands

All commands are available via the Command Palette (Ctrl+Shift+P / Cmd+Shift+P):

• Claudemeter: Fetch Claude Usage Now - Manually fetch current usage data
• Claudemeter: Open Claude Settings Page - Open claude.ai/settings in your default browser
• Claudemeter: Start New Claude Code Session - Start a new token tracking session
• Claudemeter: Show Debug Output - Open debug output channel
• Claudemeter: Log into Claude Code - Run claude auth login in a terminal (or point you at the install docs if Claude Code isn't installed). Only needed when you're not already logged into Claude Code.
• Claudemeter: Dump State (for bug reports) - Print a redacted snapshot of current state (identity, resolved org, token presence, live sessions) to an output channel - attach this to issues to speed up diagnosis

Troubleshooting

No usage shown / "Not logged into Claude Code"

• Claudemeter reads Claude Code's login. If Claude Code isn't logged in, run Claudemeter: Log into Claude Code (or claude auth login in a terminal), or just use Claude Code once.
• If Claude Code isn't installed, install it first - see claude.com/product/claude-code.
• If you authenticate Claude Code with ANTHROPIC_API_KEY (or Bedrock / Vertex / Foundry) instead of a subscription login, there's no subscription usage to show - the local Tk context gauge still works.

Usage looks like the wrong account

Claudemeter tracks whichever account Claude Code is on. If it looks wrong, check claude auth status - that's the account claudemeter reads. If you run Claude Code with a non-default CLAUDE_CONFIG_DIR, claudemeter follows that too.

Fetch errors

• Run Claudemeter: Fetch Claude Usage Now from the Command Palette.
• Run Claudemeter: Dump State (for bug reports) and report an issue with the output attached.

Reporting a bug with a log

If the steps above don't sort it, a log makes diagnosis much quicker:

1. Turn logging on - set claudemeter.debug to true in Settings (optionally set claudemeter.debugLogFile to a file path).
2. Reproduce the problem.
3. Grab the log - run Claudemeter: Dump State (for bug reports) for a short redacted snapshot, and Claudemeter: Show Debug Output for the fuller channel.
4. Open an issue and attach it, with your VS Code and extension versions. The log helps a lot. Include it.

Safe to share by design. Claudemeter redacts its logs automatically - no token, no credentials, no email, no account name, and your home directory (username) is stripped from paths. Only your own project folder names remain. Read it before you attach it and cut anything you would rather not share. See Privacy & Security below.

Privacy & Security

• No credentials stored: Claudemeter reads Claude Code's existing OAuth token and never stores, copies, or transmits it anywhere except as the Authorization header to api.anthropic.com. It never writes to Claude Code's credential store.
• No data transmission: Usage data stays on your machine.
• Zero runtime dependencies: The extension bundles no third-party runtime packages - no browser, no Chromium, no scraping stack.
• Shareable logs by design: Dump State and the debug log carry only token presence/source/scopes/expiry (never the token), plan/tier, usage percentages, counts and timestamps. No account name, no email, and the home directory (username) is stripped from paths. Your own project folder names remain. Redaction is automatic, the final check is yours - skim it before you post.
• Open source: All code is available for review.

Feedback & Issues

If you encounter any issues or have suggestions:

1. Check the troubleshooting section above
2. Review open issues on GitHub
3. Submit a new issue with:
   • VS Code version
   • Extension version
   • Error messages from the Output panel (View > Output > Claudemeter - Token Monitor)
   • Steps to reproduce

Authors

Paying it forward by the hoopy froods at HyperI (formerly HyperSec) https://hyperi.io

License

MIT License - See LICENSE file for details.

Originally inspired by claude-usage-monitor from Mark Campbell.

---

Note: This is an unofficial extension and is not affiliated with Anthropic or Claude.ai.