token-efficient

ingpoc's avatarfrom ingpoc

Use when processing 50+ items, analyzing CSV/log files, executing code in sandbox, or searching for tools. Load for data processing tasks. Achieves 98%+ token savings via in-sandbox execution, progressive disclosure, and pagination. Supports heredocs for multi-line bash.

⭐5stars🔀0forks📁View on GitHub🕐Updated Jan 2, 2026⬇️2downloads

When & Why to Use This Skill

The Token-Efficient MCP is a high-performance Claude skill designed to optimize context window usage by offloading heavy data processing to a secure sandbox environment. It enables the analysis of large CSV files, extensive system logs, and complex datasets while achieving over 98% token savings through advanced techniques like progressive disclosure, pagination, and in-sandbox code execution (Python, Bash, and Node.js).

Use Cases

  • Large-scale Data Analysis: Filtering and aggregating CSV files with 10,000+ rows by processing data in a sandbox and returning only the summarized results.
  • Efficient Log Auditing: Searching through massive application logs for specific error patterns or warnings without loading the raw text into the conversation context.
  • Resource-Intensive Scripting: Executing complex multi-line Bash or Python scripts for data transformation and batch processing while maintaining minimal token overhead.
  • Dynamic Tool Discovery: Reducing prompt bloat by searching for and loading specific tools on-demand rather than pre-loading all available tool definitions.
nametoken-efficient
description"Use when processing 50+ items, analyzing CSV/log files, executing code in sandbox, or searching for tools. Load for data processing tasks. Achieves 98%+ token savings via in-sandbox execution, progressive disclosure, and pagination. Supports heredocs for multi-line bash."
keywordscsv, logs, execute, sandbox, pagination, batch, filter, aggregate

Token-Efficient MCP

Process data in sandbox, return only results. 98%+ token savings.

Instructions

  1. Search for tool: search_tools("csv") (95% savings vs loading all)
  2. Process data in sandbox (never load raw to context)
  3. Use pagination for large files: offset, limit
  4. Return summaries, not raw data

Quick Commands

# Process CSV with filter
scripts/process-csv.sh data.csv "price > 100"

# Search logs for errors
scripts/search-logs.sh app.log "ERROR|WARN"

# Execute code in sandbox
scripts/run-sandbox.sh script.py

Token Savings

Tool Savings Example
execute_code 98%+ Run Python without loading
process_csv 99% 10K rows → 100 results
process_logs 95% 100K lines → 500 matches
search_tools 95% Find tools on-demand

MCP Tools

Tool Purpose
execute_code Python/Bash/Node in sandbox
process_csv Filter, aggregate CSV
process_logs Pattern match logs
search_tools Find tools by keyword
batch_process_csv Multiple CSVs at once

References

File Load When
references/patterns.md Choosing which tool
references/examples.md Need code examples