Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Claude Code Token Meter

    1

    Audit exact Claude Code token usage, identify startup context waste, and install a live usage meter.

    $12

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned

    Claude Code Token Meter

    Example session with this skill installed

    Show me my last 5 Claude Code sessions and tell me where the tokens are going.

    • Read your context and instructions
    • Compiled the claude code token

    I read your last 5 Claude Code sessions from ~/.claude/projects. These are exact figures from the usage data Claude Code stores, not estimates.

    Session token report

    SessionStartedFresh inCache rdOutCache hitFirst prompt
    e3fe8ac900:1755.2K1.3M9.5K96%long working session
    d41a068700:1037.7K283.1K5.8K88%file summary task
    292fb78c00:0824.8K31.3K8356%"hi"
    3ea9334c00:0714.6K29.7K8567%"hi"
    f6f7661a23:2513.4K30.9K6170%"hi"

    What stands out

    Startup cost is your biggest avoidable expense. Three sessions that only said "hi" still used 13.4K to 24.8K fresh input tokens before any work began.
    2.

    Connectors explain the gap. The 24.8K session loaded claude.ai connectors. The 13.4K and 14.6K sessions did not. That is about 10K to 11K tokens saved per session, just by starting without connectors you do not need.
    3.

    Your long session is healthy. e3fe8ac9 re-read 1.3M tokens from the cache at a 96% hit rate. Its cost comes from length, not waste. The low hit rates on the "hi" sessions are normal for sessions that tiny.

    Next steps

    1. Run the startup audit to see which files and servers fill your context: python3 scripts/audit_startup.py
    2. Type /context in a fresh session for the exact breakdown.
    3. Prove the fix with an A/B test: run the same task 3 times each way, then python3 scripts/ab_compare.py --a id1,id2,id3 --b id4,id5,id6

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Stop guessing why your Claude Code sessions are expensive. This skill audits your local session transcripts and project configuration to provide exact token usage data and actionable ways to reduce costs.

    What it does

    • Token Auditing reads local Claude Code session transcripts to extract exact input, output, and cache hit metrics.
    • Startup Analysis estimates the fixed token cost of your project's CLAUDE.md, MCP configurations, and skill front matter.
    • Live Status Bar installs a custom status line in Claude Code to monitor model usage and context pressure in real-time.
    • A/B Performance Testing compares two sets of sessions to prove if a prompt or configuration change actually saves tokens.
    • Cache Hit Monitoring identifies sessions with low cache efficiency to help troubleshoot frequent context resets.

    How it works

    1. Run a session report to view recent costs and identify the largest token consumers in your history.
    2. Audit the project startup cost to see which files are filling your context before you even send a prompt.
    3. Perform an A/B test by running a repeatable task under two different configurations to measure the delta.
    4. Optionally install the live status meter to keep token counts visible at the bottom of your terminal.

    Frameworks & tools

    Uses Python 3 scripts to process local JSON data in ~/.claude/projects. No external dependencies or network calls required.

    Why this beats prompting it yourself

    Claude cannot see its own underlying metadata or local session files through standard prompts. This skill bridges that gap by running local analysis scripts that read the actual telemetry data stored on your machine.

    Use cases

    • Identify which MCP tools or system files are bloated and wasting startup context.
    • Verify if a new CLAUDE.md instruction reduced your average session cost.
    • Debug why a session's cache hit rate is below 50%.
    • Monitor real-time context usage to avoid hitting model limits during long refactors.

    Known limitations

    Startup audits are estimates based on disk files. For exact startup breakdowns, you must manually run the /context command and provide the output. Status line installation requires manual approval.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Startup audits are estimates based on file size. Status line installation modifies settings.json. Requires Python 3. Does not read raw source code or credentials.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 8 days ago

    Software
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Environment Variables

    Reads Claude Code session transcripts and a few config files to report token usage and estimate startup cost. Reads only names (never values) from settings.json. Does not read ~/.claude.json, credentials or source code. Makes no network calls and runs no installs or deletions. Writes: a small cache file in the system temp folder (status line), and ~/.claude/settings.json only if the user runs install_statusline.py with --apply, after saving a backup copy. The optional TOKEN_METER_WARN environment variable sets the status line warning threshold.

    Listed8 days ago

    What's inside

    Frequently Asked Questions

    Nowhere. The scripts read files on your own machine and print results in your terminal. They make no network calls and send nothing to Agensi or any other server.

    Claude Code with its default session transcripts in ~/.claude/projects, and Python 3.9 or newer. There is nothing to install. The live status line needs a Claude Code version that supports the statusLine setting.

    Session totals are exact. They come from the usage fields Claude Code stores with every response. The startup audit is an estimate from files on disk (about 4 characters per token). Type /context inside Claude Code for the exact startup breakdown.

    Yes. The A/B comparison takes two groups of sessions, for example 3 runs with your old CLAUDE.md and 3 with the new one. It shows averages and ranges for each group, and warns when the ranges overlap and the difference may be noise. You must still check that the task succeeded, because fewer tokens with a worse result is not a saving.

    No. It shows tokens. On Pro and Max plans the limit that matters is your plan usage limit, so token counts are best used to compare setups with each other, not to predict a bill.

    New in LLM Ops