Serveur MCP usagebar

usagebar

Your Claude Code plan quota in the macOS menu bar, plus an MCP tool so agents can check the quota and their own context size before starting big work. An agent can decide to wait for a reset, split work, or stop before it hits the limit.

Not affiliated with Anthropic. It reads what claude -p /usage prints and the transcripts Claude Code keeps in ~/.claude/projects/. Neither format is documented, so a Claude Code release can break parsing until usagebar is updated.

Install

Homebrew

brew install dnlbit/tap/usagebar
brew services start usagebar
claude mcp add usagebar -s user -- "$(brew --prefix)/opt/usagebar/bin/usagebar-mcp"

brew services start starts the menu bar app now and at login. Update with brew upgrade usagebar: the in-app updater is off for Homebrew installs. Needs claude on your PATH (~/.local/bin, /opt/homebrew/bin or /usr/local/bin).

Uninstall:

brew services stop usagebar
brew uninstall usagebar
claude mcp remove usagebar -s user

Install script

curl -fsSL https://usagebar.dnlbit.com/install.sh | sh

Installs into ~/.local/share/usagebar (Python venv with rumps), starts at login, and prints the claude mcp add command for the MCP tool. Needs macOS, python3 3.10+ and claude on your PATH.

Updates: the app checks GitHub Releases once a day and shows "Update to vX.Y.Z" in its menu. Click it to install and restart. Nothing updates without that click. Rerunning the install command also updates.

Uninstall:

launchctl bootout gui/$(id -u)/com.dnlbit.usagebar
rm -rf ~/.local/share/usagebar ~/Library/LaunchAgents/com.dnlbit.usagebar.plist \
  ~/.config/usagebar.json ~/.cache/usagebar
claude mcp remove usagebar -s user

Menu bar

The bar shows the current session quota (▰▰▱▱▱ 48%), then the open Claude Code sessions counted per state: working, your turn, blocked (▰▰▱▱▱ 48% ●1 ○2 ⚠1). States with no sessions are not shown (▰▰▱▱▱ 48% ●1 ○2). Hover for the reset time, click for each quota with "in 1h 52m" after its reset time.

Plans without quota lines in /usage (enterprise) show the session counts only (●1 ○2 ⚠1). With nothing to count (no quotas and no counted sessions) the bar shows ◉.

Sessions

The menu lists each open Claude Code session, newest change first: <state> <name> <context>% context <status>.

• ● working: Claude is working or running a tool.
• ○ your turn: the turn is finished and the session waits for your next message. "your turn (background shell)": a background shell command still runs, but you can type.
• ⚠ blocked: the session needs you, for example a permission prompt (the row shows why).
• · a status this version does not know; the raw value shows. It is not counted in the bar.

Click a session to bring its terminal to the front. In Terminal.app and iTerm2 its tab is selected (macOS asks once for Automation permission). In other apps (Ghostty, VS Code, Cursor, Warp…) the app only comes to the front.

Session data comes from ~/.claude/sessions/*.json and the transcripts in ~/.claude/projects. Claude Code does not document these files, so a Claude Code update can change or break this part.

Settings… holds:

• Idle interval (default 20 min): poll rate while usage is flat.
• Active interval (default 5 min): used after a poll sees the percentages change, until a poll sees no change.
• Auto start at login: writes ~/Library/LaunchAgents/com.dnlbit.usagebar.plist (unchecking removes it). Takes effect at next login. Homebrew installs have no checkbox: use brew services start usagebar to start at login.
• Notify: blocked (permission prompt) (default on): a notification when a session becomes blocked, for example on a permission prompt.
• Notify: your turn (finished) (default on): a notification when a session finishes a turn that ran for 2 seconds or more and waits for your reply.

Settings are saved to ~/.config/usagebar.json. Log: /tmp/usagebar.log.

MCP tool

mcp_server.py (standard library only) gives agents one tool, get_usage, returning:

• plan: quota percentages and reset times, shared with the menu bar app through ~/.cache/usagebar/last.json. If that is older than max_age_minutes (default 2; 0 forces it), the tool runs claude -p /usage itself (about 2s), so it works without the app. fetched_at says how fresh it is.
• context: tokens in the calling Claude Code session's context, the model's max window (hard-coded table in monitor.py) and percent used.

Either part comes back as {"error": ...} on its own if it fails. Example (text, the raw /usage output, shortened):

{
 "plan": {
  "fetched_at": "2026-10-13T22:53:05-07:00",
  "error": null,
  "quotas": {
   "Current session": {"percent_used": 41, "resets_at": "2026-10-13T23:59:00-07:00"},
   "Current week (all models)": {"percent_used": 16, "resets_at": "2026-10-20T15:59:00-07:00"},
   "Current week (Fable)": {"percent_used": 0, "resets_at": "2026-10-20T16:00:00-07:00"}
  },
  "text": "You are currently using your subscription to power your Claude Code usage\n\nCurrent session: 41% used …"
 },
 "context": {
  "session_id": "a659f724-dd01-40c3-bcbe-6a3d33bd43a6",
  "model": "claude-opus-5-5",
  "context_tokens": 135210,
  "window_tokens": 1000000,
  "percent_used": 13.5
 }
}

Let your agents budget themselves

Paste this into your project or user CLAUDE.md (after the MCP setup above):

## Usage budget
- Before long or parallel work, and about every hour during it, call the
  `mcp__usagebar__get_usage` tool.
- If the session quota is above 80%, finish the current step, then stop or wait
  for the reset time it reports.
- If context is above 50%, compact or hand off to a fresh session.

Privacy

Everything stays on your machine. The only network requests are the ones claude makes itself and the daily update check to api.github.com.

Development

python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
./run.sh
python3 monitor.py --test && python3 mcp_server.py --test

A git checkout never offers updates. Release: bump __version__ in monitor.py, commit, then gh release create vX.Y.Z --generate-notes. Tags are never moved or re-pointed; fix a bad release with a new version.