Stop Hitting Claude Code Usage Limits: Best Tools and Settings (2026)
Your 5-hour Claude Code window burns out in 30 minutes? Here is what actually consumes it and the verified tools that stretch it.
If you are on the Claude Pro plan, you have probably watched a 5-hour usage window disappear in under an hour. Community threads in July 2026 are full of the same complaint. Most of the burn is not your prompts. It is context you did not know you were paying for.
This guide has three parts: how the limits actually work, free habits that cut usage, and open-source tools we verified on GitHub before recommending.
How Claude Code limits work
Anthropic documents this at code.claude.com/docs/en/costs:
- You have two windows: a rolling 5-hour window and a weekly window. Both are shared across all models and across claude.ai, Claude Code, and Claude Desktop. Switching models does not reset them.
- The full conversation is re-sent with every request. Every tool call is another request that carries the whole history. A one-line question in a session that has been open all day still bills for the entire conversation.
- Prompt caching softens this, but the cache lasts one hour on a subscription. Come back after lunch and your first message reprocesses everything.
/compactis itself a large request — it must read everything it summarizes./clearcosts nothing.
That one mechanism explains most “I barely did anything” burn: long sessions, not hard questions.
Free habits that cut the most usage
- Run
/clearbetween unrelated tasks. It is the single biggest free lever. Use/renamefirst if you want to/resumelater. - Plan in one session, implement in a fresh one. Use plan mode to produce a plan, save it to a file,
/clear, then say “implement PLAN.md”. The implementation runs on a lean context. - Keep CLAUDE.md under ~200 lines. It loads on every session and bills even when irrelevant. Move reference material into skills, which load only on demand.
- Check your effort and thinking settings. Thinking tokens bill as output. Use
/effortor lower the thinking budget for routine work. - Pick the model for the job. Sonnet for most coding, the big model only for architecture and hard debugging, Haiku for subagent grunt work.
- Run
/usageand/contextbefore you guess./usagenow shows which skills, subagents, and MCP servers consume your week, as percentages. Fix what the data names. - Point at files, do not paste them. Large pastes and vague prompts (“improve this codebase”) trigger broad, expensive scanning.
- Disable MCP servers you are not using with
/mcp. Tool definitions are deferred by default now, so this matters less than it used to — but zero is still cheaper than deferred.
Verified tools and repos
We checked each repo page before listing it. Tools split into three jobs. Only the second and third actually extend a session — a meter alone does not save tokens.
Measure where tokens go
| Tool | What it does |
|---|---|
| ccusage | npx ccusage@latest. Reads Claude Code’s local usage logs and produces daily, weekly, and per-session token and cost reports. The standard first step. |
| Claude Code Usage Monitor | Live terminal monitor with burn-rate analytics and session forecasting — predicts when you hit the limit, which is the real Pro-plan question. |
Compress what enters context
| Tool | What it does |
|---|---|
| RTK | brew install rtk. A Rust binary that sits between the shell and the agent and compresses output from 100+ dev commands (git, npm, docker, pytest) before it enters context. Claims 60–90% reduction on command output. Noisy tool output is the cheapest thing to throw away, so this is the most direct win on this list. |
| Context Mode | MCP server that sandboxes large tool output into local SQLite; the agent gets summaries and references instead of full dumps. |
| token-optimizer | Claude Code plugin with hooks that compress command output and turn repeat file reads into delta-only reads. |
Avoid reads entirely with an index
| Tool | What it does |
|---|---|
| Graphify | Builds a local, queryable knowledge graph of your codebase with tree-sitter — no API key, nothing leaves your machine. The agent queries the graph instead of re-reading files. Treat the big savings claims with skepticism: gains scale with project size, and small projects may see little benefit. |
| claude-mem | Compresses each session into semantic summaries injected into later sessions, cutting the “re-explain the project” tax at session start. |
| claude-context | Semantic code search MCP so the agent retrieves snippets instead of loading directories. Real setup cost: needs an embedding key and a vector DB endpoint. |
One honest note: viral posts assign precise savings percentages to many of these tools. Those numbers come from aggregator blogs, not the repos. The mechanisms are real; the exact percentages are marketing.
The 10-minute setup
npx ccusage@latest— see where last week actually went.- Run
/usagein Claude Code — see which skills and MCP servers eat your window. - Cut CLAUDE.md to the essentials; move the rest into skills.
brew install rtkif your sessions are command-heavy.- Adopt the two habits that need no install:
/clearbetween tasks, and plan-then-implement in separate sessions.
Do those five and the same Pro plan does several times the work.
On the free ChatGPT/Codex side, the app also grants periodic limit resets most people never redeem:
Need OpenClaw fixed live?
Remote rescue sessions for gateway, auth, tunnel, VPS, and model access problems.
See Rescue Session