Free tools Windows power users keep installed
One-click scans. No signup required.
If Claude Code usage is climbing, check whether the cause is conversation context, extended thinking, MCP tools, or additional agent requests. Anthropic documents all four as ways token use can grow; the commands /usage and /context help you investigate before changing your workflow.
Start by finding what is using tokens
In Claude Code, run /usage to check current token usage and /context to see what is occupying the conversation context. These answer different questions: usage helps you inspect the session’s consumption, while context helps identify material being carried into or added to it. Anthropic’s cost guide recommends both: Manage costs effectively.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
On Pro, Max, Team, and Enterprise plans, Anthropic’s cost guide describes an additional recent-usage breakdown that can attribute activity to skills, subagents, plugins, and individual MCP servers, with behavior flags such as long context and cache misses. Treat those figures as approximate: they are calculated from local session history on that machine and do not include activity on other devices or Claude.ai. Attribution details can also depend on the Claude Code version, so check your installed version before relying on a particular display.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →For an organization-wide view, Claude Code’s monitoring documentation describes the claude_code.token.usage and claude_code.cost.usage metrics, with dimensions that can include token type, user, team, model, skill, plugin, or agent. Anthropic cautions that “Cost metrics are approximations.” Use your API provider’s billing records, such as Claude Console or the applicable cloud provider, as the billing source of truth. See Anthropic’s monitoring documentation.
#1 Best Overall
Compare like with like
Token totals can be misleading if cache categories are handled differently. Claude Code’s exported input count excludes cache reads and writes unless the corresponding cache-token fields are added; monitoring reports cache reads and cache creation separately. Check which categories a dashboard includes before comparing its total with another usage display or a bill.
1. Old conversation context follows you into new work
Anthropic explains that “Token costs scale with context size: the more context Claude processes, the more tokens you use.” When a conversation grows, later requests may continue processing earlier context, including details that no longer matter to the task. This can make an unrelated follow-up cost more than starting with a clean conversation.
Clear when the task changes
Use /clear when moving to unrelated work. It starts a fresh conversation rather than carrying old task context forward. Anthropic’s cost guide summarizes this advice as “Clear between tasks.”
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Compact only when continuity is useful
/compact summarizes earlier conversation so work can continue with a shorter context. It is not free: Claude must read and summarize the prior conversation. When you need continuity, provide custom compaction instructions that retain only the decisions, constraints, or facts required for the next step. When you do not need that history, clearing is the simpler option.
2. Extended thinking can add output tokens
Extended thinking is useful for tasks that benefit from more deliberate reasoning, but Anthropic bills thinking tokens as output tokens. The cost guide says the default thinking budget can reach tens of thousands of tokens per request, depending on the model. That does not mean every request uses that amount; it means a task that does not need extended reasoning may be using a capability without a corresponding benefit.
For straightforward tasks, choose a lower effort level or disable thinking if the current model and task support that choice. Model behavior differs across releases: some models always use extended thinking, while others offer controls. Check the documentation for the model and version you actually use instead of assuming one setting applies everywhere. Anthropic’s current guidance is in its Claude Code cost guide.
3. MCP definitions and tool results take context
Model Context Protocol (MCP) integrations can add context through tool definitions and the results returned by tools. Anthropic says definitions are deferred by default, so the mere presence of a configured server does not imply the same fixed token overhead in every setup. Tool output matters too: verbose results can add substantial context even when the definition itself is not currently loaded.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute- Run
/contextto see whether tool-related material is occupying context. - Run
/mcpto review configured MCP servers. - Disable servers you are not actively using, and avoid requesting unnecessary detail from tools.
- Where a suitable CLI command is available and more context-efficient, use it instead of an MCP tool call.
These are workflow choices, not hardware upgrades. Keep an integration when its convenience or capability is worth the context it adds; remove or limit the ones that are not helping with the current task.
4. Subagents and agent teams add their own requests
Delegating work can keep detailed intermediate steps out of the main conversation, but it does not make the delegated requests disappear. Subagents make their own requests, and agent teams run separate instances, each with its own context window. Anthropic says team token usage scales with active teammates and runtime.
Anthropic also gives an approximate comparison: agent teams can use about 7× more tokens than standard sessions when teammates run in plan mode. This is a qualified figure for that team configuration, not a universal multiplier for every subagent, team, or task.
Keep delegation proportional to the task
- Delegate only work that benefits from parallel execution or a genuinely separate investigation.
- Keep agent prompts focused so each worker has less irrelevant context to process.
- Use a lower-cost model for simple delegated work when appropriate and available.
- Keep teams small and stop teammates when their work is complete.
The trade-off is parallelism versus additional requests: delegating can reduce elapsed time or keep the main conversation concise, but may increase total token use.
Recommended Free Tools
Check whether local usage matches billing
Local usage views, exported metrics, and provider bills can differ because they have different scopes and accounting. Local cost attribution uses session history on the current machine; monitoring cost metrics are approximate; and cache reads and writes may be separate from the input total. Reconcile equivalent categories and use the API provider’s billing records when you need the billed amount.
Best Value
A gateway can affect attribution and billing as well. Anthropic documents that gateway credentials may route requests on a per-token basis to the credential owner, and requests made with a gateway credential may not count toward subscription usage limits. Anthropic also states that it does not endorse, maintain, or audit third-party gateways. If you use one, confirm which credential is making the request and where its usage is billed in Anthropic’s gateway documentation.
Choose the smallest change that addresses the cause
- Long context: clear unrelated work; compact only when you need to preserve selected information.
- Thinking: lower effort or disable it when the model and task permit.
- MCP: remove inactive servers and limit unnecessarily verbose tool results.
- Agents: delegate selectively, focus prompts, and stop workers when finished.
Start with /usage and /context, then change the workflow indicated by the evidence. The goal is not to minimize every request; it is to avoid spending tokens on context, reasoning, integrations, or parallel work that the task does not need.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
