Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Context Rent: What Eats Your Token Budget

    1

    On a real 192-turn Claude Code session, **96% of all tokens were context being re-read**, and output was under 1%.

    Free

    Security scanned
    Context Rent: What Eats Your Token Budget

    Context Rent: What Eats Your Token Budget

    Example session with this skill installed

    I keep hitting my Claude Code usage limit by mid-afternoon. What is eating my tokens in this project, and what should I change?

    • Read your context and instructions
    • Compiled the context rent

    Summary

    This is the real report from the Claude Code session that built a Dart payments package (192 turns).

    96% of all tokens were context being re-read; output was under 1%. Three causes account for most of it.

    Where the tokens went

    tokensshare
    Context re-read from cache49.1M96%
    Context written to cache1.7M3%
    Output (incl. thinking)129k<1%

    What filled the context (size × turns it stayed)

    Sourcecallsrawcarried
    screenshots11833k4.0M
    shell output4615k1.1M
    other MCP tools769.1k1.1M
    web pages4215k836k
    file reads22.0k306k

    Top causes

    1. One long session with a large context (~17M tokens). 192 turns, average context ~265k. Every turn re-sent all of it.
    → One task per session; run /compact at a natural break instead of waiting for auto-compaction (790k → 29k tokens, once).

    2. 36 screenshots stayed in context (~1.8M tokens). Browser screenshots were ~19k tokens raw but were carried for 2.7M.
    → Read pages as text (read_page / get_page_text) and take one screenshot only to check layout, at a reduced scale.

    3. The prompt cache expired 6 times during breaks (~1.4M tokens). After pauses of 2 to 12 hours, ~230k tokens of context were written again each time.
    → After a long break on a big context, run /compact or start a fresh session with a short summary.

    Suggested CLAUDE.md section

    
    ## Context budget
    - Prefer reading pages as text; take a screenshot only to verify layout, at a reduced scale.
    

    Next steps

    Shall I add this section to CLAUDE.md? For the next big task, I suggest starting a new session and compacting at each milestone.

    Suggested CLAUDE.md

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

    About this skill

    Why your usage limit runs out

    An agent sends its whole context again on every turn. On a real 192-turn Claude Code session, 96% of all tokens were context being re-read, and output was under 1%. So what matters is what sits in your context, and for how long.

    A 2,000-token screenshot that stays for 100 turns costs 200,000 tokens. Context Rent charges every tool result for exactly that (size × turns it stayed in context) and turns the biggest rent-payers into fixes.

    What it finds

    FindingFix it suggests
    Long sessions with a big contextOne task per session, /compact at task boundaries
    Screenshots kept in contextRead pages as text, one screenshot per batch
    Whole-file readsGrep first, read a range
    The same file read 3+ timesEdit from what is already in context
    Noisy test/build/install outputQuiet flags, | tail, with the exact command named
    Prompt cache expired during breaks/compact or start fresh after a long pause
    Large CLAUDE.md / AGENTS.mdKeep rules, move reference material out
    MCP servers configured but never usedRemove them
    Auto-compactionCompact earlier

    It also writes a ready-to-paste Context budget section for CLAUDE.md / AGENTS.md.

    How the agent uses it

    1. Runs the bundled zero-dependency Node script on your local logs (~/.claude/projects, ~/.codex/sessions).
    2. Explains the shape of your usage and the top three causes, with numbers.
    3. Applies only the fixes you approve: CLAUDE.md lines, a trimmed instruction file, removing unused MCP servers.

    Privacy

    Runs offline. Reports contain sizes and short labels only (a file path, the start of a command), with API keys and tokens masked. Your prompts, code and command output never appear in the output.

    Open source (MIT): https://github.com/MBotirjonDev/context-rent

    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.

    Requires local access to Claude Code or Codex CLI session logs. Estimates sizes at ~4 chars per token. Read-only until user approves config edits.

    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

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 9 days ago

    • Free with an Agensi account

    Software
    Git

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Listed9 days ago
    Updated2 days ago

    What's inside

    Frequently Asked Questions

    Your agent re-sends its whole context on every turn, so a tool result pays "rent" for every turn it stays in context. Context Rent reads your local Claude Code and Codex CLI logs, charges each tool result for size × turns kept, ranks the biggest causes (long sessions, screenshots, whole-file reads, noisy commands, cache expiry, large CLAUDE.md, unused MCP servers) and applies only the fixes you approve.

    This skill is specifically designed for local CLI-based agents, including Claude Code and Codex CLI, by interfacing with their local log directories. It is not compatible with web-browser interfaces like Claude.ai or ChatGPT.

    The skill folder: SKILL.md (the workflow your agent follows) and three zero-dependency Node.js scripts (rent.cjs, parse.cjs, analyze.cjs) that read your logs and produce the text, JSON or Markdown report. The same code is open source (MIT) on GitHub.

    No. It runs offline on your machine and makes no network calls. Reports contain only sizes and short labels such as a file path or the start of a command, with API keys and tokens masked. Your prompts, code and command output are never printed.

    Compacting shrinks one session after the fact. Context Rent tells you what filled the context in the first place, across all your sessions, so you can stop it: which commands, files, screenshots and instruction files cost the most, and when the cache expired during breaks. It also tells you when compacting earlier would have helped.

    New in LLM Ops

    Free