See which tasks consume provider capacity.

Subscriptions have limits, and you need to know which work spends them. Agent Tank, the open-source (MIT) usage tracker built by the ProPR team, ships inside ProPR's agent image and shows the capacity each Claude Code, Codex, and Antigravity task consumed.

Open Agent Tank website →

Agent Tank in the sidebar: Claude session, weekly and Fable windows, the Codex weekly window, and Antigravity's weekly and five-hour limits for Gemini and for Claude and GPT, each as a usage bar — Agent Tank in the sidebar: every Claude Code, Codex and Antigravity window at a glance.

One run on Claude Opus 5.5: 19 minutes, 2.4M tokens in, and 2.4% of the Claude session and 0.6% of the weekly window

Usage per task

ProPR measures provider usage before and after each task and charges the difference to that task.

Every window in one view

The usage windows of Claude Code, Codex, and Antigravity sit side by side. Agent Tank covers these three agents.

Agent Tank in the sidebar, collapsed: one bar each for the Claude session, the Codex weekly window and the Antigravity Gemini weekly limit

Live sidebar

Current usage and reset countdowns update while work runs.

History per job and per window

Each agent call carries its usage record, so the task history shows which job spent the capacity. ProPR keeps 15 days of readings: hover a usage row for its last two periods, or chart every agent's windows in Analytics.

LLM log rows tied to their work: two pull requests, a plan and a repository summary, each with its tokens, cost and the capacity its call used

Agent Tank quota usage over the last seven days: the Claude session window filling and resetting through each working day, and the weekly windows of Claude, Codex and Antigravity climbing toward their resets — Seven days of subscription capacity, per agent and window.

Turn it on in Settings.

Tracking is off by default. Bundled (recommended) runs Agent Tank on demand inside ProPR's agent image, with nothing to install and no daemon. External points ProPR at a daemon you run.

Unattended runs pause near the limit

Scheduled Automations read the latest capacity snapshot and pause at 90% by default. Per-call cost lives in observability, and the pull request cost guide puts capacity and per-call spend side by side.

Agent Tank setting with Disabled, Bundled and External modes; Bundled is selected and ready — Agent Tank's integration modes in ProPR Settings.