Agent Usage Monitor

A VS Code editor tab showing how much of your AI coding-agent allowance is left, for Codex, Claude Code and Antigravity (Gemini) side by side.

Everything is read from files those tools already keep on your machine, plus the same usage endpoints their own clients call, with the credentials they already stored. There is no server, no account, and no telemetry — nothing leaves your machine except the usage requests to each vendor.

Limits at a glance

What it shows

Limits at a glance

One card per tool, showing how much of the current window is left and when it resets. The bar turns yellow past 40% used and red past 80%, and the countdown is the thing you actually read — "in 57m" matters more than the timestamp. Codex adds its credit balance and available full resets; Claude adds extra-usage spend for the month; Antigravity reports a rolling quota shared by all its Gemini models, and publishes no weekly figure.

Tool	Limits shown	Source
Codex	5-hour + weekly, credit balance, plan	`chatgpt.com/backend-api/wham/usage`, falling back to the last session log
Claude Code	5-hour + weekly, extra-usage spend, plan	`api.anthropic.com/api/oauth/usage`
Antigravity	rolling quota per model pool (no weekly figure is published)	`cloudcode-pa.googleapis.com/v1internal:fetchAvailableModels`

Token usage

Totals per tool for the last 24 hours, 7 days or 30 days, then a full breakdown: fresh input, cache reads and writes, output, and how much of that output was reasoning. Cache reads usually dwarf everything else, which is worth seeing.

Token usage

Daily chart and projects

The three tools side by side per day, with a data-table view for exact numbers, and which of your projects the tokens went to.

Daily tokens and projects

Models

Which models did the work, per tool.

Models

Install

From the VS Code Marketplace, or search Agent Usage Monitor in the Extensions view.

Or build it from source:

git clone https://github.com/jiajilah/agent-usage-monitor.git
cd agent-usage-monitor
python3 build-vsix.py
code --install-extension jiajilah.agent-usage-monitor-0.1.1.vsix

Reload the window, then run Agent Usage: Open Usage Monitor from the Command Palette (Ctrl/Cmd+Shift+P).

Over Remote-SSH, install it on the remote side — that is where the agents' data lives.

Requirements

VS Code 1.96 or newer
python3 on the machine the extension runs on (it decodes Antigravity's SQLite + protobuf data)
Whichever agents you use, logged in. Each tool is optional; missing ones are reported, not fatal.

Settings

Setting	Default	Meaning
`agentUsage.lookbackDays`	`90`	How far back to scan session logs
`agentUsage.codexHome`	`~/.codex`	Override the Codex home directory
`agentUsage.claudeHome`	`~/.claude`	Override the Claude Code home directory
`agentUsage.geminiHome`	`~/.gemini/antigravity`	Override the Antigravity data directory
`agentUsage.antigravityModel`	`gemini`	Model-name prefix whose Antigravity quota is shown. All `gemini-*` models share one allowance, so the default covers every Gemini version

Where the numbers come from

Codex — token counts from ~/.codex/sessions/**/*.jsonl; live limits using ~/.codex/auth.json. Session logs also record rate_limits, so the card falls back to the last logged values if the live call fails.
Claude Code — token counts from ~/.claude/projects/**/*.jsonl; live limits using ~/.claude/.credentials.json. Claude never writes limit percentages to disk, so there is no fallback; its money fields are in cents.
Antigravity — token counts decoded by antigravity_usage.py from the per-conversation SQLite databases in ~/.gemini/antigravity/conversations/*.db. Each step carries a protobuf blob whose usage block was identified by cross-checking two tables that record the same calls; output tokens equal thinking + response on every call observed. Conversations saved in the older *.pb format are encrypted and cannot be counted. Live quota uses ~/.gemini/jetski-standalone-oauth-token.

Parsed log files are cached by size and modification time in the extension's global storage, so only the first scan is slow. While the tab is open it refreshes every 5 minutes.

Caveats worth knowing

These are undocumented, internal endpoints. They are the ones each vendor's own client calls, but they can change without notice and break this extension. The Antigravity request must send a User-Agent identifying the Antigravity client or the API returns 403.
Antigravity's field names are inferred. Its data is raw protobuf with no field names; the labels here come from cross-checking, not documentation.
Antigravity publishes no weekly limit. That does not prove no weekly cap exists — only that none is reported.
Antigravity omits a zero quota. When its allowance runs out, remainingFraction disappears from the response rather than reading 0; a missing value is treated as exhausted.

License

MIT — see LICENSE.