A local-first macOS menu bar app for tracking AI coding token usage, estimated cost, and relay account capacity.
VibeToken is built for developers who use AI coding tools throughout the day and want one quick, honest view of local usage without opening multiple dashboards.
Note
This is an early preview with twenty-two supported AI coding sources. The current release is v0.1.4 Preview. Downloadable builds are available from GitHub Releases, but they are not yet Apple-notarized.
- Automatic usage collection from 22 AI coding sources, with no simulated increments.
- Today, rolling 24-hour, 7-day, and 30-day totals and trends.
- Input, cache, output, reasoning, model, tool, and session breakdowns.
- Versioned OpenAI, Anthropic, and Google API price estimates with pricing coverage shown explicitly.
- Duplicate emission and fork/subagent replay filtering.
- Optional read-only Sub2API pool monitoring with per-account remaining quota, plan expiration dates when provided, account-specific estimated recovery times, and explicit rate-limited or unavailable states for Codex 5-hour and 7-day windows.
- Native macOS menu bar and Dock entry points in English and Simplified Chinese; Dock visibility and launch at login are managed only in the dedicated Settings window, which opens without dismissing the menu bar popover.
| Source | Status |
|---|---|
| Codex Desktop / CLI | Supported: live and archived sessions, including local profiles |
| Claude Code + Claude Desktop Code/Cowork | Supported: project logs, profiles, and Cowork session roots |
| Gemini CLI | Supported: current JSONL, legacy JSON, and nested subagent sessions |
| OpenCode | Supported: read-only SQLite with legacy JSON fallback |
| GitHub Copilot CLI | Supported: exact shutdown model metrics |
| Cursor | Supported: official account usage export, refreshed at most every 5 minutes |
| Cline | Supported: standalone and VS Code-family task logs with migrated-copy deduplication |
| Roo Code | Supported: VS Code-family task logs, including current and legacy indexes |
| Kiro CLI | Supported: native session logs; Token totals are explicitly estimated from observed text |
| Grok Build TUI / CLI | Supported: exact per-turn usage and per-model splits |
| DimAgent | Supported: read-only usage ledger with fork replay deduplication |
| OpenClaw | Supported: agent sessions, named profiles, and legacy data roots |
| pi | Supported: exact assistant-message usage, including cache reads and writes |
| Qwen Code | Supported: Gemini-style usage metadata with exclusive cache and reasoning categories |
| Kimi Code | Supported: current agent wires and legacy Kimi session stores |
| MiMoCode | Supported: read-only SQLite; imported external sessions are excluded |
| Amp | Supported: usage ledger with message-level fallback |
| Droid | Supported: exact cumulative totals; time distribution is derived from observed growth |
| Hermes | Supported: default and named-profile SQLite stores |
| Trae CLI | Supported: trace-level model usage with span deduplication |
| Antigravity | Supported: current offline protobuf databases and legacy local language-server fallback |
| ZCode | Supported: read-only message usage database |
| Sub2API Codex account pool | Supported, optional |
VibeToken does not infer exact token usage from ChatGPT or Claude desktop conversations. Subscription usage and API-priced cost are different things, so estimated cost is always labeled as an estimate.
The current GitHub Release build requires macOS 14+ on Apple silicon. On launch, VibeToken automatically discovers every supported source. There is no source setup, folder picker, or manual scan step. Missing tools are skipped without blocking the sources that are available.
- Open the latest release, then download the
macos-arm64VibeTokenarchive and the checksum file. - Extract the archive and move
VibeToken.appto Applications. - On first launch, Control-click the app in Finder and choose Open. This preview uses an ad-hoc signature and is not Apple-notarized, so macOS may show a security warning. Download builds only from this repository's Releases page.
Building from source also requires Xcode 16 or Swift 6 command-line tools:
git clone https://github.com/AaAndrew233/VibeToken.git
cd VibeToken
swift test
./scripts/build-app.sh
open "dist/VibeToken.app"The build script creates an ad-hoc signed app at dist/VibeToken.app. It is suitable for local use, but it is not equivalent to a Developer ID signed and notarized release.
- Current release: v0.1.4 Preview.
- For end-user distribution, use the versioned archive and checksum published on GitHub Releases. Do not share the entire development folder.
- The current archive is ad-hoc signed and not notarized, so another Mac may show a Gatekeeper warning.
- A public end-user release should use Developer ID signing, Apple notarization, and a versioned archive or DMG.
- Open VibeToken and click its menu bar item.
- Choose Today, 24H, 7D, or 30D.
- Select live, 5-minute, 30-minute, or manual refresh for local usage collection. A connected Sub2API pool checks account membership every 30 seconds and performs a verified quota refresh every 30 minutes.
- Click the gear button in the top-right of the menu bar popover to open Settings and control Dock visibility and launch at login. The popover remains open while Settings appears; these two options are not duplicated elsewhere in the popover.
- Use the language control to switch between English and Simplified Chinese.
For optional Sub2API monitoring, sign in with an administrator account in the existing Relay Capacity area. After the first sync, each detected Plus account uses Plus (1x), while every detected Pro account must be assigned Pro 5x, Pro 10x, or Pro 20x manually. An unconfigured Pro account does not receive a guessed default, and pool capacity remains unavailable until it is configured. App startup, Mac wake, a user-initiated refresh, a changed account pool, and the 30-minute schedule all request Sub2API's official usage probe with force=true. On released servers without the batch route, VibeToken falls back to the released per-account endpoint with source=active&force=true and at most six concurrent requests. It then reads the account list back until every active physical account has a newly persisted, valid 5-hour and 7-day snapshot, or an explicit unavailable or exhausted state. The pool total is published only after the entire set passes validation. A partial or unverifiable refresh hides the current total, reports the verified account count, and keeps only the timestamp of the last successful refresh. VibeToken does not reset, edit, or delete relay accounts.
Token totals come from structured usage fields. Kiro CLI is the exception: its native session log has no Token counters, so VibeToken labels its text-based estimate accordingly. Cost is estimated from known model prices:
estimated cost = input * input price
+ cache write * input price
+ cache read * cache price
+ (output + reasoning) * output price
Unknown models remain unpriced instead of receiving a guessed fallback price. Historical usage is currently recalculated with the price catalog bundled in the installed app.
The bundled catalog records its verification date, effective dates, and official OpenAI, Anthropic, and Google source URLs. A usage snapshot selects time-limited pricing from its latest event timestamp; for example, Claude Sonnet 5 switches from its introductory price on September 1, 2026. A range that crosses a price-change boundary therefore remains an aggregate estimate rather than an invoice-grade event-by-event calculation. OpenCode reuses the matching provider price when its model identifier is recognized.
Cache writes use the normal input price. The current estimator does not apply per-request long-context premiums or Gemini cache-storage time because aggregated local logs do not preserve those billing dimensions reliably. Subscription plans, free tiers, provider discounts, taxes, and tool-call charges are also excluded. Token usage is still counted when a model has no matching bundled price; the UI marks the cost as partial or unavailable instead of inventing a rate.
Pricing sources: OpenAI, Anthropic, and Google Gemini.
For Sub2API, physical account counts remain unweighted. The availability card and plan summary show currently available / schedulable total, for example Plus 7/8 · Pro 4/4, without mixing runtime-unavailable accounts into that total. The separate quota-capacity percentage keeps all active accounts in its denominator so unavailable capacity is not hidden. Capacity is weighted as Plus = 1, Pro 5x = 5, Pro 10x = 10, and Pro 20x = 20, then normalized to 100%. Each account contributes the smaller remaining value of its 5-hour and 7-day windows. Temporarily unavailable, explicitly rate-limited, exhausted, stale, or unobserved accounts remain in the total capacity denominator but contribute zero currently available capacity. The account field shows the plan expiration date below the account name when the server provides it. The account quota sheet is split into Available Accounts and Unavailable Accounts tabs, defaults to Available Accounts, and shows each tab's count. Only accounts with a normal runtime state, current quota data, and remaining capacity in both windows are considered available. The 5-hour and 7-day values remain on one line, and estimated recovery appears below them only when the account is marked rate-limited and every active blocker has a valid future reset time. Accounts are sorted with valid Pro accounts first, earlier recovery times first within a plan, valid Plus accounts next, and invalid accounts last. The pool-level estimate is the earliest complete recovery among those accounts. Shadow accounts are excluded.
- Only structured usage, model, project, session, and timestamp fields are extracted. Conversation content is never stored or sent.
- Source JSON/JSONL files and SQLite databases are opened read-only.
- Cursor is the only usage source that requires a provider request: VibeToken reads the existing Cursor access token from
state.vscdbin memory and sends it only tohttps://cursor.comto fetch the official usage CSV. The token and response body are not persisted or logged. - Legacy Antigravity
.pbhistory is decoded only through its already-running language server on127.0.0.1; its CSRF token is kept in memory and never logged. Current Antigravity.dbhistory is read offline. - Usage indexes stay in the local application support directory.
- Tokens, passwords, cookies, account addresses, and response bodies are excluded from logs.
- Sub2API credentials are stored in local files with restricted permissions, not in macOS Keychain. This is less protected than Keychain against other processes running as the same macOS user.
VibeToken is an independent Swift implementation. It has no runtime dependency on another usage collector: each adapter reads the supported tool's structured local data or documented account export directly.
swift build
swift testSee CONTRIBUTING.md before changing a data source, usage formula, persistence model, or security boundary.