Claude Code as a personal OS: which weekly quota bucket your routines actually drain

It is the Context-Connections-Capabilities-Cadence pattern Anthropic and a handful of writers (MindStudio, Towards AI) have been pushing since the Routines feature shipped on 2026-04-14. Context is what Claude knows about you (CLAUDE.md, memory files). Connections is data it can reach (MCP, Google Workspace, your repo). Capabilities is skills it can run (markdown files in .claude/skills/). Cadence is when it acts without you (remote routines on Anthropic's cloud, scheduled tasks, Auto Mode loops, Cowork). When people say 'personal OS' they almost always mean the Cadence step. That is the step that drains weekly quota differently from interactive Claude Code, and the step every other guide stops short of explaining.

Two buckets, simultaneously. seven_day_oauth_apps because the routine authenticates as an OAuth client (same as your local Claude Code CLI), and seven_day_cowork because cloud-executed routine work is itself classified as Cowork on the server. Both fields appear verbatim in claude-meter's UsageResponse struct (models.rs lines 24 and 26). The chat UI on claude.ai aggregates everything into seven_day; the cowork and oauth_apps buckets are not shown there. That is why your account can sit at seven_day = 30 percent and seven_day_cowork = 100 percent at the same time, and why a scheduled routine fires '429 rate_limit_error' while a web chat in another tab still works.

Pro is 5 remote routine runs per day, Max is 15, Team and Enterprise are 25. Minimum scheduled interval is one hour. Multiply by seven: Pro caps at 35 runs per week, Max at 105. That is the daily-cap math; the weekly quota is independent and stricter. Every routine run also charges your regular plan usage the same way an interactive Claude Code session would, so a five-run-per-day Pro account that is also coding interactively will hit seven_day_oauth_apps or seven_day_cowork well before the 35-run quota wall. Source: code.claude.com/docs/en/routines.

ccusage reads the JSONL files Claude Code writes to your local machine when you run it interactively. Remote routines run on Anthropic's cloud (4 vCPUs, 16GB RAM, 30GB disk per run, per Anthropic's docs) against a cloned copy of your GitHub repo. Nothing about that run touches your laptop, so nothing about that run lands in ccusage's data source. The token estimate you see locally stays low. The seven_day_cowork bucket on the server fills up. They are measuring different things; ccusage is not wrong, it is just blind to this surface. ClaudeMeter polls api.anthropic.com/api/oauth/usage every 60 seconds, which returns the same shape the chat UI's settings page calls, so the cowork bucket shows up live.

It is a flat JSON object with one Window field per bucket. Each Window has utilization (a float, where 1.0 is the wall) and resets_at (a UTC ISO-8601 timestamp, rolling, not calendar). The fields you care about for a personal OS setup are five_hour, seven_day, seven_day_oauth_apps, seven_day_cowork, and extra_usage. claude-meter's parser names every field at models.rs lines 19-28, which is the open-source documentation for an endpoint Anthropic does not document publicly anywhere else.

Sometimes. If extra_usage.is_enabled is true and used_credits is below monthly_credit_limit on /overage_spend_limit, prompts that would have walled bill against extra usage at API prices instead. If is_enabled is false, or out_of_credits is true, the wall stays. Anthropic's April 2026 metered-billing rollout makes this the realistic escape hatch for a personal-OS setup that genuinely needs to keep running, but you want the live dollar burn visible. ClaudeMeter renders extra_usage as a third row in the popover when it is present.

Yes, but slightly differently. /loop is a local skill, so your laptop has to stay open and your session active. The traffic still goes through your OAuth-authenticated Claude Code CLI, so seven_day_oauth_apps still increments. Whether seven_day_cowork increments depends on whether the work classifies as Cowork on the server side; in practice for any agentic loop where Claude is figuring out the steps itself, it does. The local visibility is better (ccusage sees it now, because the traffic is local), but the weekly buckets you blow through are the same two.

Pull the access token Claude Code stores in macOS Keychain (`security find-generic-password -s 'Claude Code-credentials' -w`), then curl the endpoint: `curl -H 'Authorization: Bearer $TOKEN' https://api.anthropic.com/api/oauth/usage`. The response is the exact UsageResponse JSON shape claude-meter parses. The OAuth token has user:profile scope, which is sufficient. If the response is missing seven_day_cowork entirely, your account does not have a Cowork bucket yet (rollouts are staged); seven_day_oauth_apps is the one to watch in that case.