U.S.M.C. MinistriesU.S.M.C. MinistriesFleet Command

Usage & Progress

Mission Control for frontier-model allowance + work finished. Generated 2026-10-29T11:44:58-04:00.

Jump to: who spent it ↓

Most consumer frontier plans do not expose remaining quota via clean API. This board shows LIVE measured OpenClaw usage + manual/browser snapshots for remaining allowance.
Edit ~/.openclaw/shared-memory/context/fleet-usage-manual.json then re-run fleet-usage-progress-build.py
⚠ Snapshot stale (>48h): Cursor, Gemini / AI Studio, NotebookLM. Nightly scrape runs 03:40 ET or edit fleet-usage-manual.json.

Standing rules

  • Comfort band is 80–90% used. Crossing 80% is enough. 85–90 is fine. Do not grind the week to 100%.
  • Two failure modes were named: running out about three days early, or climbing the last two days from about 40% to 80%. That is the pattern to avoid. It is not today's percent.
  • A reply that dies at Sir is an output or context limit. It is not a quota hit.
  • SuperGrok Heavy annual is under consideration. It is not bought. Do not purchase from this board. Hold the stated October expiry until Adam says buy.
  • Plan names on the cards are labels, not receipts. ChatGPT tier is not confirmed here. Claude: the fleet runs on Adam's Claude Max account; its week/5-hour percentages on the card come live from the Anthropic usage API (weekly reset Wed ~15:00 ET).
  • Grok and ChatGPT percentages on the cards come live from the OAuth sign-ins OpenClaw already holds (`openclaw models status`). The Heavy meter is grok.com. console.x.ai is a different meter.
  • Scorecard refine and commerce enrichment are pointed at grok-4.7, not local Qwen. Pastor extract still points at local 1235, then 1234. Do not flatten those lanes.
  • Hermes on this machine defaults to Codex gpt-5.6-sol (fallbacks Claude Opus 5, grok-4.7, local Qwen). Max and nine desks run Claude Opus 5.5; Prof X, Rush, SgtMaj Mac and Nimble run grok-4.7. Fable is capped this week. Do not retune from this page.
  • A login wall, a bot check, or a block page is not remaining quota.
  • Pacing (Adam, 2026-10-23): Claude week should sit at 80–90% by Tuesday afternoon and about 90% by Wednesday noon, never dry on Monday morning. A meter near empty within 24h of its reset is the plan, not an alarm.

Spoken 2026-09-26. Re-checked against this machine on 2026-09-29. Model/Hermes/Claude/Grok lines corrected 2026-10-14 against openclaw.json, ~/.hermes/config.yaml and Max's sprint plan. Not a meter reading. Claude/Grok/ChatGPT lines updated 2026-10-23 when the live poll (frontier-usage-poll.py) replaced browser snapshots.

14/17
Fleet ops
2
Wins today
24
Wins 7d
41
Cowork projects
98
Enabled jobs
434
Wins archive

Provider usage

Grok / SuperGrok (xAI)
SuperGrok (OAuth) · manual_snapshot · snapshot 0h ago
check
66% remaining
Weekly 34% used (resets in 5d 9h)
102runs 24h
0tokens 24h*
222runs 7d
Weekly percent comes live from the SuperGrok OAuth sign-in OpenClaw holds (openclaw models status). console.x.ai is a different meter.
top: grok-4.7×102
ChatGPT / Codex (OpenAI)
ChatGPT / Codex (OAuth) · manual_snapshot · snapshot 0h ago
check
69% remaining
Week 31% used (resets in 5d 20h) · 5h 25% used (resets in 8m)
59runs 24h
0tokens 24h*
158runs 7d
Week and 5-hour percents come live from the Codex OAuth sign-in OpenClaw holds (openclaw models status). Tier on this card is not a receipt.
top: gpt-6-astra×58, gpt-5.6-sol×1
Claude Max / Claude Code
Claude Max · manual_snapshot · snapshot 0h ago
check
92% remaining
Week 8% used (resets Wed 13:59 ET) · 5h 14% used (resets Thu 13:59 ET) · pace 13% expected now → on pace (-5 pts)
233runs 24h
0tokens 24h*
610runs 7d
Week and 5-hour percents come live from the Anthropic usage API behind Claude Code's sign-in. Pace line targets 85% by Tue 15:00 ET; weekly reset Wed ~15:00 ET.
local 7d: 0 rows · 13,688,319 output tokens
top: claude-opus-5-5×218, claude-haiku-4-5×7, claude-sonnet-5×5
Cursor
Cursor Pro/Pro+/Ultra · needs_snapshot · STALE 968h
check
remaining: not snapshotted yet
0runs 24h
0tokens 24h*
0runs 7d
Remaining budget only on Cursor dashboard. Update manual snapshot after checking Usage.
top: —
Gemini / AI Studio
AI Studio / API · measured_usage_only · STALE 968h
check
remaining: not snapshotted yet
0runs 24h
0tokens 24h*
0runs 7d
RPM/RPD caps; remaining not a clean counter. Measured Gemini runs shown.
top: —
NotebookLM
NotebookLM · needs_snapshot · STALE 968h
check
remaining: not snapshotted yet
0runs 24h
0tokens 24h*
0runs 7d
No public remaining-query API. Snapshot notebooks/sources/daily hits manually.
top: —
Local LM Studio
local · healthy
check
remaining: not snapshotted yet
0runs 24h
0tokens 24h*
0runs 7d
Local. No subscription ceiling. Not the scorecard or commerce grind brain.
top: —

* Tokens only when provider reports them into OpenClaw logs.

Who spent it

Which agent spent the Claude allowance, and on what. The Claude card above is the real meter; this section measures what used it. Tap any agent to drill in. Loading…

Where it went (whole fleet)

    Which Claude model

    Agents, ranked

    Bar length = share of the fleet. Bar colors = where it came from (same colors as above). Right side = % of fleet and approximate meter points.

      How to read this

      • The Claude card above is the real meter. It comes live from Claude's own usage reading. This section is our measurement of what used it.
      • Points are estimates. One meter point is about 2.8 "weighted Mtok", give or take 50%. Trust the ranking and the shares. Treat the point numbers as rough.
      • Context size is what burns the meter, not how much an agent writes. A long-running chat re-reads its whole history on every turn.
      • Agent-to-agent messages are charged to the agent that receives them. A drill-down shows who sent them, and what each agent causes on others.
      • Scheduled jobs on Max include "relays": a job owned by another agent whose output Max then had to process.
      • Heartbeats run on the local model, so they barely show. Grok and GPT do not count against Claude; their call counts are listed for context only.
      Method and caveats

      • Source: the Claude Code session logs on the fleet host Mac. Each API call is counted once (deduplicated by message id).
      • Agent = the workspace the call ran in. Source = the message that opened that turn.
      • Weights: input 1, cache write 1.25, cache read 0.1, output 5; Sonnet ×0.6, Haiku ×0.2 (API price ratios).
      • Not measured: claude.ai web/mobile chats and Claude Desktop chats (not on disk). Grok reports no token counts, so it shows calls only.
      • Counts only. No message text, prompts or chat ids are published.

      OpenClaw model volume (24h)

      ProviderModelRunsTokensOK/Bad
      anthropicclaude-opus-5-52180217/1
      xaigrok-4.71020102/0
      openaigpt-6-astra58057/1
      anthropicclaude-haiku-4-5707/0
      anthropicclaude-sonnet-5505/0
      anthropicclaude-fable-5-1303/0
      openaigpt-5.6-sol101/0

      Latest wins

      DateAgentTitle
      2026-10-29popsTACC Morning Brief — Thu Oct 29
      2026-10-29popsEmail back-brief (2026-10-29)
      2026-10-28sallyBrad's Mac setup SOP (step by step, with a check per step)
      2026-10-28popsThe WATCH Bill: household duties on one page. Draft for Sunday's huddle.
      2026-10-28popsThe SWO's Vision: five fronts on one page, ready to print.
      2026-10-28popsTACC Morning Brief — Wed Oct 28
      2026-10-28popsEmail back-brief (2026-10-28)
      2026-10-27sallyBrad and Adam summary (fixed: 2 pages, no blank space)
      2026-10-27sallyBrad and Adam: summary, action items, bottom line
      2026-10-27popsTACC Morning Brief — Tue Oct 27