More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Agent Cost Scan — What Every Run Really Costs

    1

    Reads your agent logs and reports cost per run, the model eating the budget, and three ranked fixes with the money each one recovers.

    Free

    Security scanned
    Agent Cost Scan — What Every Run Really Costs

    Agent Cost Scan — What Every Run Really Costs

    Example session with this skill installed

    Our support agent's bill went from about $200 to $520 a month and nobody can tell me why. The logs are in logs/agent-*.jsonl — one line per model call.

    Work out what a run actually costs, find the thing that's out of proportion, and tell me what to fix first. We budget $0.40 per run in CI, so also tell me if we're over.

    • Read your context and instructions
    • Compiled the agent cost scan

    Summary

    Scanned

    4,482 calls across 260 runs over 12 days. Spend is

    $167.31, a run-rate of

    $422.87/month. Cost per run is

    $0.64 mean, $0.55 median, $1.28 p95 — over a $0.40 budget at the tail, so this would fail CI today.

    MetricValue
    Total spend, 12 days$167.31
    Monthly run-rate$422.87
    Cost per run$0.64 mean / $1.28 p95
    Prompt cache55% hit, $150.99 already saved
    Retry waste3.8% of spend, $6.42

    The one thing

    claude-opus-5-5 is 34% of spend on 11% of calls. Nothing else in the log is that far out of proportion.

    ModelCostShareCalls
    claude-sonnet-5-5$109.8965.7%2,918
    claude-opus-5-5$56.2433.6%479
    claude-haiku-5-5$1.180.7%1,085

    Fixes, ranked by money recovered

    Route routine calls off claude-opus-5-5 — $78.17/month. Keep it for the hard step; send classification, lookup and summarising to Haiku 5.5.
    2.

    Raise prompt cache hit rate — $35.87/month. Hit rate is 55%, target 70%. Put system prompt, tool schemas and policy text in one stable prefix and never edit it mid-run.
    3.

    Stop paying twice for retries — $11.35/month. 3.8% of spend is retried calls, mostly schema failures. Validate the response before the next attempt, not after.

    Next steps

    • Move the three classification calls to Haiku 5.5 and re-run the scan on the next 7 days.
    • Add --budget-per-run 0.40 to CI so this cannot regress silently.
    • Run it again after the change; the report is the before/after.

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

    About this skill

    Your agent has a bill. It doesn't have a number.

    Agent Cost Scan reads the logs you already have and tells you what one run costs — then shows where the money goes and what to fix first. Three minutes, no API key, no dashboard to connect.

    What you get

    • Cost per run — mean, median and p95. A single monthly total hides the runs that actually hurt.
    • Monthly run-rate — extrapolated from the period you logged, so the number is comparable to an invoice.
    • The one thing — the model whose share of spend is far above its share of calls. Usually the whole story.
    • Cache and retry waste — what caching already saves you, and what failed calls cost twice.
    • Three ranked fixes — each with a stated mechanism and the monthly money it recovers.
    • A budget gate — --budget-per-run 0.40 exits 1 when p95 crosses the line, so a prompt change that doubles cost fails CI instead of arriving on the invoice.

    Priced on this year's models, not last year's

    The table ships with list prices checked on 13 November 2026 — Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.2, the GPT-6 family and Gemini 4 Argon — including per-model cache rates and long-context tiers. Cache reads are not a flat tenth any more: Opus 5.5 reads at $0.20 against $4.00 input, and a tool that assumes otherwise overstates your saving.

    Works with what you have

    Anthropic and OpenAI SDK usage records map directly, and any JSONL can be shaped to the schema. Python 3.10+, standard library only. It reads local files and does arithmetic — it never calls a model.

    Why it is not another cost dashboard

    Dashboards give you a total. This gives you a number you can defend in a budget conversation: cost per run, a named culprit, and three levers with figures attached. The report says "recoverable", never "saved", and every assumption is written down in SKILL.md.

    Includes

    • SKILL.md your agent loads directly
    • scripts/cost_scan.py — the scanner
    • A synthetic 4,482-call support-agent log to try it on before you trust it with yours
    • The report and JSON generated from that log
    • A self test covering parsing, published numbers, bad input and exit codes

    Good to know

    Figures are estimates from list prices in the included table — override them for your contract or fine-tune. The scanner reads usage metadata only and ignores prompt text, but strip sensitive fields before sharing a log.

    Requires JSONL logs with token counts. Estimates based on list prices, not custom contracts. Does not analyze prompt content or latency.

    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 any payment.

    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 forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    • Free with an Agensi account

    Requires

    Python 3.10+
    A JSONL log of your agent model calls

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Listed1 day ago

    What's inside

    Frequently Asked Questions

    The skill identifies 'waste' by detecting inefficient retry loops, high-cost model calls that don't yield proportional utility, and poor prompt cache hit rates that result in unnecessary charges.

    It is a standalone Python tool compatible with any agent framework (like LangChain, AutoGen, or CrewAI) as long as you can export your execution logs to a standard JSONL format.

    You receive the core Python analysis script, a collection of pre-configured pricing tables for major providers (OpenAI, Anthropic, Gemini), and documentation for CI/CD integration.

    While it comes with built-in pricing tables updated for late 2026, the tool supports custom JSON price tables so you can manually adjust for your specific enterprise discounts or new model releases.

    Setup is minimal: ensure you have Python 3.10+ installed and point the script at your JSONL log directory to generate your first audit report.

    No, this skill is a post-execution audit tool; it analyzes logs from past runs to identify trends and regressions without adding latency to your live agent calls.

    New in LLM Ops