More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    LLM Spend Advisor

    1

    Turn local LLM usage records into clear spend audits, cost comparisons, and budget decisions with llm-usage MCP.

    $2.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    LLM Spend Advisor

    LLM Spend Advisor

    Example session with this skill installed

    Audit my LLM API spend for the last 7 days and compare it with the preceding 7 days. Break down the results by provider, model, and project, explain what drove any increase by separating call volume from average cost per call, and identify the biggest savings opportunity.

    Then project next month’s cost for 10,000 calls averaging 6,000 input tokens and 1,500 output tokens per call. I have a $100 budget—compare suitable priced models, show which ones fit, and recommend the cheapest option. Use recorded data only, state the exact UTC time windows and assumptions, exclude failed or partial calls by default, and do not change my model configuration.

    • Read your context and instructions
    • Compiled the llm spend advisor

    Recorded spend increased from $1 to $11 (+1,000%). The main driver was higher recorded cost per call in demo / gpt-5-mini. For next month’s workload, Qwen qwen-turbo is the cheapest recorded-price option at $6, leaving $94 of your budget.

    All figures below use local records only. No model configuration was changed.

    Audit windows, start-inclusive and end-exclusive:

    • Previous: 2026-10-12T09:14:33Z → 2026-10-19T09:14:33Z
    • Recent: 2026-10-19T09:14:33Z → 2026-10-26T09:14:33Z

    Both are exactly seven days. Failed and partial calls are excluded.

    BreakdownPrevious calls → recentPrevious spend → recentIncrease
    Total1 → 3$1 → $11+$10
    Provider: OpenAI1 → 2$1 → $9+$8
    Provider: Anthropic0 → 1$0 → $2+$2
    Model: gpt-5-mini1 → 2$1 → $9+$8
    Model: claude-sonnet-4-60 → 1$0 → $2+$2
    Project: demo1 → 2$1 → $9+$8
    Project: unassigned¹0 → 1$0 → $2+$2

    ¹ Derived from total minus assigned projects. Provider, model, and project rows are separate views of the same spend; do not add them together. Growth percentages for zero-baseline groups are undefined.

    What drove the increase

    Calls increased 200%, while average recorded cost per call rose from $1 to $3.6667 (+266.7%).

    Using prior-period cost per call as the baseline:

    • Volume effect: (3 − 1) × $1 = +$2
    • Average-cost effect: 3 × ($3.6667 − $1) = +$8

    These sum to the $10 increase; this convention assigns the interaction to average cost. Within demo / gpt-5-mini, the $8 increase separates into $1 from volume and $7 from average cost. Anthropic contributes another $2 of new activity.

    Average recorded tokens stayed at 1,000 input and 100 output per call in both periods. Longer average prompts or responses therefore do not explain the increase in these aggregates.

    Biggest savings opportunity—and a data discrepancy

    Prioritize investigating demo / gpt-5-mini: it accounts for 81.8% of recent spend and 80% of the increase. Its two recent calls recorded $9, but their combined 2,000 input / 200 output tokens would cost only $0.0009 at the stored ordinary input/output rates. The Anthropic tokens similarly imply $0.0045 versus $2 recorded.

    These are material discrepancies, not proven recoverable savings. Historical pricing or recording details remain unresolved; I have preserved the recorded totals rather than repricing them.

    For a prospective model change, the largest savings among your observed models would be replacing a full Sonnet workload with qwen-turbo: $405 → $6, saving $399 (98.5%), conditional on task suitability. Against gpt-5-mini, the same workload saves $39 (86.7%).

    Next month’s projection

    Assumed month: 2026-11-01T00:00:00Z → 2026-12-01T00:00:00Z. Your 10,000 calls imply 60 million input and 15 million output tokens, with a $0.01-per-call budget.

    Selected priced candidates across all four recorded providers:

    Provider / modelInput / output $ per millionCost/callMonthly costFits $100?
    Qwen qwen-turbo$0.05 / $0.20$0.000600$6.00Yes
    OpenAI gpt-5-nano$0.05 / $0.40$0.000900$9.00Yes
    Qwen qwen-flash$0.05 / $0.40$0.000900$9.00Yes
    OpenAI gpt-4.1-nano$0.10 / $0.40$0.001200$12.00Yes
    DeepSeek deepseek-v4-flash$0.14 / $0.28$0.001260$12.60Yes
    OpenAI gpt-4o-mini$0.15 / $0.60$0.001800$18.00Yes
    DeepSeek deepseek-chat$0.28 / $0.42$0.002310$23.10Yes
    Anthropic claude-3-haiku-20240307$0.25 / $1.25$0.003375$33.75Yes
    OpenAI gpt-5-mini$0.25 / $2.00$0.004500$45.00Yes
    Anthropic claude-haiku-4-5$1.00 / $5.00$0.013500$135.00No
    Anthropic claude-sonnet-4-6$3.00 / $15.00$0.040500$405.00No

    Recommend qwen-turbo on price, subject to validating quality, context capacity, tool support, latency, and availability. The records rank costs; they do not establish those suitability requirements.

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

    About this skill

    High LLM API bills often hide behind untracked workloads and missing project attribution. This skill audits recorded usage, identifies cost spikes, and projects future spend using the llm-usage MCP server. It transforms local database records into evidence-based spend reports without manually exported CSVs or provider dashboard hopping.

    What it does

    • Spend auditing analyzes recorded API calls by provider, model, or project across specific ISO-8601 time windows.
    • Cost driver analysis isolates whether spend increases stem from rising call volume or changes in average cost per call.
    • Workload projection calculates estimated costs for future deployments based on expected token counts and volume.
    • Model comparison ranks providers and models by price for specific workloads to find the most cost-effective candidates.
    • Budget planning derives per-call allowances from remaining monthly budgets and identifies models that fit a stated budget.

    How it works

    1. Scope the request by resolving the time window, project filters, or workload assumptions provided in the prompt.
    2. Execute MCP queries using usage_summary for headlines or query_spend for detailed provider and model breakdowns.
    3. Analyze and project by comparing windows or calculating estimates via compare_providers and get_pricing.
    4. Deliver the report as a compact table detailing USD cost, call counts, and actionable savings opportunities.

    Frameworks & tools

    This skill requires an active connection to the llm-usage MCP server. It interacts with local SQLite databases storing captured API events and uses the bundled pricing catalog for cost calculations.

    Why this beats prompting it yourself

    Generic prompts cannot access your local usage history or bundled pricing snapshot. This skill understands the specific schemas of the llm-usage server, ensuring accurate window filtering and preventing mathematical errors when comparing multi-tag events or partial-stream records.

    Use cases

    • Audit weekly spend to find which project or developer is driving the highest API costs.
    • Compare candidate models (eg. GPT-4o versus Claude 3.5 Sonnet) costs for a new 100k-call-per-month workload.
    • Investigate why costs doubled overnight by comparing volume and per-call averages.
    • Set a per-call budget and find which models fit the limit based on 2k-token prompts.

    Known limitations

    Does not track flat-rate chat subscriptions like ChatGPT Plus. Failed or partial-stream calls are excluded by default. Rankings are based on price only, not model quality or latency.

    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 a separately installed local llm-usage-mcp MCP server. Does not track flat-rate subscriptions or provider invoices. Rankings are based on price only, not output quality, latency, context capacity, or task suitability.
    Failed and partial-stream calls are excluded by default. Projections use the installed pricing snapshot rather than live provider quotes.

    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 22 days ago

    Software
    llm-usage-mcp 0.1.3 or laterPython 3.13 or later

    Permissions

    No special permissions declared or detected

    Listed22 days ago

    What's inside

    Frequently Asked Questions

    This skill requires the `llm-usage` MCP server and a local SQLite database that records your API activity; it analyzes your historical data rather than providing a generic estimation tool.

    It specializes in identifying hidden cost drivers by comparing call volumes against model pricing changes and calculating specific 'per-call' allowances to keep your projects within a set monthly budget.

    The skill uses a bundled pricing catalog to ensure calculations are accurate; while it doesn't auto-update, you can refresh the pricing data through the connected MCP server to reflect the latest market rates.

    Yes, you can audit and project spend by filtering for specific projects, developers, or ISO-8601 time windows, provided your local usage records include those metadata tags.

    The advisor focuses on consumption-based API spend; it does not track flat-rate monthly subscriptions like ChatGPT Plus, Claude Pro, or Gemini Advanced.

    You receive the structured skill logic designed to interface with the llm-usage MCP server, enabling automated spend reporting, workload projections, and model cost comparisons.

    New in LLM Ops