Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    deepseek spend audit

    1

    Estimate DeepSeek spend from local Claude Code transcripts, with peak and off-peak pricing.

    Free

    Security scanned

    deepseek spend audit

    Example session with this skill installed

    Show me my DeepSeek spend report for the last 7 days. Why was my usage so high yesterday?

    • Read your context and instructions
    • Compiled the deepseek spend audit

    Claude Code's internal cost tracking uses Claude model rates, so this report re-calculates your actual spend using DeepSeek's specific pricing and cache discounts.

    ¥18.42 across 12 sessions in the last 7 days, 92% of it on the deepseek-reasoner model.

    DeepSeek Spend Report (Last 7 Days)

    ModelPeak TokensOff-Peak TokensCache HitsTotal Cost (CNY)
    deepseek-chat1,240,000450,00088%¥1.47
    deepseek-reasoner3,800,000120,00042%¥16.95
    Total5,040,000570,000N/A¥18.42

    Findings

    1. The deepseek-reasoner model accounts for ¥16.95 of your spend. At 4.5x the cost of the chat model, moving non-reasoning tasks to deepseek-chat would cut costs significantly.
    2. Your cache hit rate on the reasoner model was only 42% yesterday. This suggests large prompts are being modified frequently, preventing you from accessing the 1/50th price discount for cached input.
    3. 96% of your reasoner tokens were consumed during peak hours (UTC 06:00-10:00). Moving these batch refactoring tasks to off-peak windows would reduce this specific line item by 50%.

    Next steps

    • Review session sess_01J2K3L4 to see why the cache hit rate dropped so low.
    • Use deepseek-chat for routine boilerplate generation to preserve reasoner credits.
    • Run large codebase indexing tasks during off-peak UTC windows.

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

    About this skill

    Claude Code records session transcripts locally, but its built-in cost tracking often fails to reflect DeepSeek's specific pricing models or cache discounts. This skill parses your local history to provide an accurate breakdown of your actual API spend.

    What it does

    • Transcript parsing scans local ~/.claude/projects logs for raw token counts.
    • DeepSeek pricing applies current peak and off-peak rates to every turn.
    • Cache analysis calculates the impact of cache-hits versus misses on your bill.
    • Model breakdown isolates spend between deepseek-v3 and deepseek-r1 models.
    • Usage reporting generates structured summaries of credit consumption by window or session.

    How it works

    1. Run the script with specific flags like --days or --session to target specific data.
    2. The tool maps model IDs and timestamps against the DeepSeek pricing table.
    3. Review the generated table showing model-specific costs and token volumes.
    4. Analyze findings regarding peak-hour usage and cache efficiency to reduce future spend.

    Frameworks & tools

    Built with Python 3.9 using only the standard library for zero-dependency execution. It specifically targets Claude Code transcript schemas.

    Why this beats prompting it yourself

    Raw prompts cannot access your hidden local filesystem where Claude Code stores history. This skill bridges the gap between your local logs and DeepSeek's complex tiered pricing, providing visibility that the Claude CLI lacks by default.

    Use cases

    • Audit weekly spend across multiple development projects.
    • Explain why a specific long-running session burned through credits.
    • Identify if switching to off-peak hours would significantly cut costs.
    • Verify cache hit rates to optimize prompt engineering strategies.

    Known limitations

    Does not account for public holidays in off-peak calculations. Only tracks turns with explicit deepseek-* model IDs; third-party proxies that mask model names will be excluded from totals.

    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 Python 3.9+. Does not account for public holidays. Only tracks deepseek-* model IDs; proxied sessions with masked IDs are ignored.

    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 5 days ago

    • Free with an Agensi account

    Software
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Listed5 days ago

    What's inside

    Frequently Asked Questions

    This skill provides granular visibility into your DeepSeek API costs by parsing hidden Claude Code logs, specifically accounting for peak/off-peak pricing and cache hits that standard CLI tools often overlook.

    The skill is designed specifically for Claude Code users utilizing DeepSeek-V3 or DeepSeek-R1 models and is compatible with any system running Python 3.9 or higher.

    You will receive the Python-based audit script, a README with execution flags, and documentation on how to map your local directory for immediate use.

    Yes, the script is a zero-dependency tool that runs locally on your machine, ensuring your private session transcripts are never uploaded to an external server for analysis.

    The tool targets specific DeepSeek model IDs; if you use a third-party proxy that renames the model (e.g., 'my-custom-proxy-1'), the script will not recognize it without manual configuration.

    New in LLM Ops

    Free