AI usage and cost, in your macOS menu bar.
Subscription limits for Claude, OpenAI and Gemini, plus ccusage-style cost estimates and real billed spend.
indicators.jonymusky.com · Download · Build from source
| Rate-limit windows (%) | Cost estimate from local logs | Real billed spend | |
|---|---|---|---|
| Claude | 5-hour session, weekly, per-model weekly, extra usage — from your Claude Code login | ~/.claude/projects transcripts |
Anthropic Admin API (optional key) |
| OpenAI | Plan windows from your Codex CLI / ChatGPT login (falls back to the last values Codex wrote to its log) | ~/.codex/sessions rollouts |
OpenAI Costs API (optional admin key) |
| Gemini | Per-model quota buckets from your Gemini CLI Google login | ~/.gemini/tmp/*/chats sessions |
— |
| Grok | Grok Bot usage pool with its reset, from your Grok Bot app login (xAI does not expose other subscription limits) | — | xAI Management API: month-to-date spend and prepaid balance (optional key + team id) |
| Cursor | Plan usage (with the Auto / named-model split) and on-demand spend for the billing cycle, from your Cursor login | — | — |
OpenCode sessions are also read and attributed to the vendor they used (Anthropic → Claude, OpenAI, Google → Gemini, xAI → Grok).
- Menu bar:
Claude 24% 5h · OpenAI 96%— the window closest to its limit for each provider (configurable: names or initials, session or weekly, optional today's cost). - Popover: every window with its reset time, today / 7-day / month-to-date cost, a 30-day spend sparkline, tokens today, per-model breakdown, and the vendor-billed spend when a key is configured.
- Notifications: optional macOS alerts when a window crosses 50/80/90/95% (your pick) and when a window resets. Each event fires once.
- Cost estimates are computed exactly like ccusage: tokens from local logs × public API list prices from LiteLLM's table (bundled snapshot, refreshed automatically). Cache writes at the 1-hour TTL rate are priced correctly. If you are on a subscription, this is what the same work would have cost via API; if you use API keys, add them under Settings → Billing APIs to see the real invoice.
brew install --cask jonymusky/tap/indicatorsThe build is ad-hoc signed, so Gatekeeper may block the first launch: right-click → Open once, or run xattr -dr com.apple.quarantine /Applications/Indicators.app.
curl -fsSL https://indicators.jonymusky.com/install.sh | shDownloads the latest release into /Applications/Indicators.app, clears the quarantine flag and launches it.
Grab Indicators.zip from the latest release, unzip, move Indicators.app to /Applications, open it.
The app is ad-hoc signed; the first launch needs a right-click → Open (or xattr -d com.apple.quarantine /Applications/Indicators.app).
Requires macOS 14+ and the Xcode Command Line Tools (xcode-select --install). No full Xcode needed.
git clone https://github.com/jonymusky/indicators.git
cd indicators
make app # → build/Indicators.app
make install # copies it to /Applicationsmake demo re-records the README media and make video renders docs/demo.mp4 (both need Pillow and ffmpeg). make cli prints the same cost report in the terminal:
Claude (Claude Code) — 338 files
today $65.98 · 7d $707.57 · 30d $7069.19 · month $4508.67
claude-fable-5-1 $65.98 in 3588 cacheRead 37547381 cacheWrite 2330787 out 198774
OpenAI (Codex CLI) — 746 files
today $9.92 · 7d $522.94 · 30d $971.59 · month $826.86
Nothing to configure for the basics: if you use Claude Code, Codex CLI or Gemini CLI, Indicators reads their local logs and reuses their logins to ask each vendor for your limits.
- Claude: sign in to Claude Code once (
claude). Indicators reads the OAuth token from the Keychain item Claude Code creates; macOS will ask you to allow access the first time. - OpenAI: sign in to Codex CLI with ChatGPT (
codex login). API-key logins have no plan limits. - Gemini: sign in to Gemini CLI with Google (
gemini→ Login with Google). API-key logins show costs but no quota buckets. - Cursor: just be signed in to the Cursor app. Indicators reads its local session and asks cursor.com for the billing cycle usage. Cursor 3.x encrypts the session, so macOS asks once to let
securityread "Cursor Safe Storage": choose Always Allow. - Grok Bot: just be signed in to the Grok Bot app. Indicators reads its local session (macOS asks once for "Grok Bot Safe Storage") and shows the usage pool on the Grok card.
The puzzle icon in the popover opens the Skills window: every skill in your agents' global folders (~/.agents/skills, ~/.claude/skills, ~/.cursor/skills, ~/.codex/skills…), where it came from, and whether its GitHub source has changed since you installed it (the same folder-hash check npx skills update runs, cached for six hours; set GITHUB_TOKEN to raise the API limit). Update, remove and add go through the skills CLI (npx skills … --global), so Node.js must be installed. Your own skills (plain folders or symlinks into your repos) are listed as Local and are never touched. The Discover tab searches the skills.sh directory and installs a skill globally in one click. From the terminal: indicators-cli skills and indicators-cli skills find <query>.
- Billing APIs (optional, Settings → Billing APIs): Anthropic admin key (
sk-ant-admin…), OpenAI admin key, xAI management key + team id. Keys are stored in your Keychain and only ever sent to the vendor that issued them.
indicators-mcp ships inside the app bundle. It is a local MCP server (stdio) that exposes the same data to any agent, plus routing rules you define:
| Tool | What it does |
|---|---|
get_usage |
Windows, % used, reset times, plan, cost estimates and billed spend per provider |
recommend_model |
For a task, returns the primary model unless its provider is at/above the threshold, then the first fallback with headroom, with the reason |
set_routing_rule / list_routing_rules / delete_routing_rule |
Manage rules in ~/.config/indicators/routing.json (task default sets the catch-all) |
estimate_cost |
USD cost of a request at list prices |
list_models |
Models seen in your logs today and which CLIs are installed |
Register it:
# Claude Code
claude mcp add indicators -- /Applications/Indicators.app/Contents/MacOS/indicators-mcp
# Codex CLI (~/.codex/config.toml)
[mcp_servers.indicators]
command = "/Applications/Indicators.app/Contents/MacOS/indicators-mcp"
# Gemini CLI (~/.gemini/settings.json)
{ "mcpServers": { "indicators": { "command": "/Applications/Indicators.app/Contents/MacOS/indicators-mcp" } } }Example rules file:
{
"defaultThreshold": 90,
"rules": [
{ "task": "code-review", "primary": "claude-opus-5", "fallback": ["gpt-6-astra", "gemini-3.1-pro-preview"], "threshold": 85 },
{ "task": "chat", "primary": "claude-sonnet-5", "fallback": ["gpt-5.4-mini"] }
],
"fallbackRule": { "task": "default", "primary": "claude-sonnet-5", "fallback": ["gpt-6-astra"] }
}The server recommends; it does not switch models by itself. It shines in orchestrators and in agents that decide which CLI or API to delegate a subtask to.
Everything runs locally. No telemetry, no backend, logs are read-only. See SECURITY.md for the exact files and endpoints.
- Antigravity (its logs are protobuf without a public schema; help welcome)
- History view: sessions per project with cost
- Notarized builds
Contributions welcome — see CONTRIBUTING.md.
Cost math and log formats follow ccusage; the vendor usage endpoints were documented by CodexBar and friends. Pricing data by LiteLLM.



