unlimited

Read AI-subscription quota usage, per window, as facts — and where each window is heading.

uv tool install unlimited                # or: pipx install unlimited

unlimited                                # usage per account, for people
unlimited read [--vendor V]... [--max-age SECONDS] --json
unlimited capture claude-statusline      # in a Claude Code statusline script

Each reading gives each limit's window length, the fraction used so far, and its reset time. It also gives whether the vendor says the limit is held, and why, and, where the vendor names it, the account's plan. unlimited reports what the vendor says, and projects it forward: it never picks an account or draws a threshold. Choosing what to do with a reading is up to the consumer.

For people

For agents

unlimited read --json gives the same readings, one per account, for load-balancing and pacing work across subscriptions (one limit shown):

[
  {
    "schema": 1,
    "vendor": "anthropic",
    "account": "work",
    "taken_at": "2026-09-21T14:00:00+00:00",
    "source": "api",
    "plan": "default_claude_max_20x",
    "status": "ok",
    "why": null,
    "retry_until": null,
    "limits": [
      {
        "name": "seven_day",
        "window_minutes": 10080,
        "used_at_least": 0.41,
        "resets_at": "2026-09-25T11:00:00+00:00",
        "held": false,
        "held_why": null,
        "severity": null,
        "active": null,
        "kind": null,
        "projection": {
          "at_reset": [0.873, 0.9436],
          "exhausts_at": null,
          "run_out": 0.164,
          "samples": 2,
          "past_windows": 5,
          "since": "2026-09-20T14:00:00+00:00"
        }
      }
    ]
  }
]

Read projection.at_reset against used_at_least rather than either alone, and past_windows/samples for how much the projection rests on.

Sources

The network endpoints are not officially documented and may change without notice. unlimited reads each harness's credential where the harness keeps it. It never refreshes a token or writes a credential, and no token appears in its output, cache or errors.

Projection

Each limit also carries a projection: where the window is heading, from the readings unlimited has taken. Within the window, two paces are extended to the reset — the average since it opened, and a recency-weighted pace with a half-life of a fourteenth of the window, which rises with a burst and falls in a quiet spell. Past windows of the same limit, kept for eight windows, add the shape of use (quiet nights, busy Mondays): each one's use from this point to its end, shifted to today's. They take over from the paces between the third and eighth window.

at_reset is [low, high], unclamped, so 1.07 means use would pass the limit. exhausts_at is when the high end reaches it, if before the reset. run_out is how likely use passes it, from the past windows that did, or null without them. samples, past_windows and since say what it rests on. A projection is only as good as how often something reads: a statusline capture feeds it continuously.

Cache

There is one cache file and one flock per vendor under $XDG_CACHE_HOME/unlimited, mode 0600. Concurrent callers share one upstream request. A 429's Retry-After holds off every caller until it passes, capped at 24 hours; a 429 or 5xx holds off at least five minutes, and the account's last good reading stands in the meantime. On macOS the keychain is readable only from the user's GUI session or launchd, not over plain ssh.

Status

Pre-release. The schema ("schema": 1) may still change.

Releasing

main takes changes only through a pull request that passes the tests and the identity check. Bump version in pyproject.toml in one, merge it, then tag the merge commit vX.Y.Z and push the tag: the Publish workflow puts it on PyPI by trusted publishing, with no stored token.

2mw2lt

unlimited is part of 2mw2lt — Too Much Work, Too Little Time — a steering partner that coordinates work across AI workers and trusted people. 2mw2lt reads its accounts through unlimited.