Why do I hit my Claude Code weekly limit so fast?

Because every request counts, and each one resends your conversation's context. Each parallel session, each subagent and each agent teammate sends its own requests. Long sessions, cache misses after a break, heavier models and anything you do on claude.ai draw from the same weekly limit. Run /usage in Claude Code to see which of these is costing you the most.

Below: how usage is counted, the usual causes (from Anthropic's own docs, as of October 2026), and what to change.

How usage is counted

Anthropic's help center lists what affects usage: the length and complexity of your conversations, the features you use, the model, and the effort level. Two facts follow from the docs and explain most surprises:

1. Claude Code sends your full conversation with every request. Each time Claude uses a tool, it sends another request carrying the tool results. Prompt caching makes re-reading that history cheaper, but it is still read. A one-line question in a session that has been open all day draws usage for the whole conversation.
2. All surfaces share one budget. Usage on claude.ai, Claude Code and Claude Desktop counts towards the same limit.

This article quotes no plan numbers. They change, and Anthropic does not publish fixed figures.

The usual causes

Anthropic's cost docs list why usage climbs in a long session. In practice, these are the ones that hit people running several agents:

• Parallel sessions. Four terminals running Claude Code are four conversations, each sending requests with its own context. They all count against the same account limits.
• Subagents and workflows. Every subagent, and every agent a dynamic workflow spawns, sends its own requests on top of the main conversation's.
• Agent teams. Each teammate is a separate Claude Code instance with its own context window, and it keeps consuming tokens until it exits. Agent teams are experimental and disabled by default (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 turns them on).
• Long context. A session you never clear gets more expensive per message.
• Cache misses. Your first message after a break longer than the cache lifetime reprocesses the full context.
• Background work. Session-scoped scheduled tasks (/loop), messages from your other sessions, and goal check-ins can start new turns while a session sits idle, each sending the full context.
• Compaction. /compact reads the conversation it summarizes, so compacting a large context is itself a large request.
• Model and effort. Opus and higher effort levels use more per request, and a switch to Opus also applies to subagents that inherit your session's model.

Step 1: find out what is using it

Run /usage. On Pro, Max, Team or Enterprise, it shows a breakdown under the plan usage bars:

• Attribution: the share of recent usage that went to skills, subagents, plugins and individual MCP servers.
• Recommendations: suggested changes, one per row, such as trimming a large CLAUDE.md.
• Loops: the heaviest /loop or other session-scoped scheduled tasks, with how often each fired and its token total.

Press d or w to switch between the last day and the last week. When the footer shows a hint such as f to fix 2 of 3, press f and Claude Code applies the recommendations it can (this needs Claude Code v2.1.305 or later). The breakdown only covers sessions on this machine. Usage from other devices or from claude.ai is not in it.

Step 2: change what costs the most

Pick from Anthropic's documented options, starting with what the breakdown points at:

/clear          # start fresh between unrelated tasks (costs nothing)
/rename         # name a session before clearing, so /resume can find it
/model          # switch model mid-session
/effort         # lower the reasoning effort for routine work
/context        # see what is filling the context window

• Clear between tasks. Stale context is paid for on every message that follows.
• Match the model to the job. Anthropic's docs say Sonnet handles most coding tasks well and costs less than Opus. For simple subagent tasks, set model: haiku in the subagent's configuration.
• Trim always-loaded instructions. CLAUDE.md is loaded into context at session start, so its tokens are there even during unrelated work. Move workflow-specific instructions into skills, which load only when invoked.
• Watch background work. Stop loops you no longer need. To stop goal check-ins, set CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0. To hold messages from other sessions instead of delivering them as new turns, set crossSessionInbound to hold.
• Keep agent teams small. If you use them, keep each teammate's task small and self-contained.
• Run fewer agents at once when the weekly bar is high. Parallelism is the multiplier on everything above.

Step 3: keep an eye on it

The weekly limit rarely surprises people who can see it. Add a status line that prints rate_limits.seven_day.used_percentage, as shown in how to check your Claude Code usage limits. For a menu bar version, see how to show Claude Code usage in the menu bar or status line.

Limits of this approach

• /usage tells you what happened, not what a new agent will cost before you start it.
• Each status line lives in one terminal and refreshes on that session's events. With many sessions open, you read the same account-wide figure in several places, at different moments.
• When you are deep in work, you don't open /usage. The check that matters most is the one you skip.

With Kanbell

Kanbell is a macOS app with one window per project and real terminals in panes and tabs. For Claude Code, it shows your subscription usage for the session and the week in the app, so you can see it while you decide whether to launch one more agent. Its agent map shows Claude Code agents and their subagents, which makes it easier to notice how many are running at once.

Kanbell does not reduce consumption. It only shows what is happening. The changes above are still what lowers your usage.