Recoleta
GitHub
Source note

Agent-stack – one command to make any repo token-efficient for Claude Code

Authors
mahdikaz
Source
Hacker News
Published
Collection
Software Intelligence
Claude CodeToken OptimizationCode IntelligenceDeveloper AgentsRepo Automation
Open GitHubMarkdown source

Summary

agent-stack is a one-command setup tool that configures Claude Code and Cursor for lower-token work in a software repo. It matters because repo agents often waste context on file discovery, noisy logs, oversized instructions, and manual hook setup.

Problem

  • Claude Code token-saving tools are split across separate utilities for shell compression, code maps, usage measurement, handoff, hooks, and editor rules.
  • Setting up a repo can require choosing 5-10 tools, merging hooks, writing CLAUDE.md, mirroring Cursor rules, and measuring usage by hand.
  • The practical cost is higher input-token use and more setup time before an agent can work on a codebase.

Approach

  • The main command, npx @drmahdikazempour/agent-stack init --all, detects the host, repo type, package manager, and profile, then generates Claude Code and Cursor files.
  • It writes and verifies CLAUDE.md, AGENTS.md, .claudeignore, skills, hooks, Cursor rules, and .agent-stack/graph.md, with backups and rollback on failure.
  • A built-in code map indexes source files and exported symbols, so the agent can grep one compact file before opening source files.
  • A built-in compress command removes ANSI codes, folds duplicate lines, and trims long command output before it enters context.
  • Usage measurement relies on ccusage; a Stop hook logs turns to .agent-stack/usage.jsonl, and measure --since 7d compares current input tokens/day with the stored baseline.

Results

  • Setup claim: init --all takes a repo from no setup to an optimized Claude Code and Cursor setup in under 2 minutes.
  • Example install output claims 20 generated files, 2 wired hooks, a verified CLAUDE.md, a generated code map, and a 7-day baseline of 12,340 tokens/day.
  • The generated CLAUDE.md is capped at ≤800 startup tokens and checked by doctor.
  • The code map example indexes 142 files and 906 top-level symbols, letting the agent open 1 target file instead of reading many files during symbol search.
  • The output compressor claims about 60% fewer characters on a 500-line log.
  • The measurement example reports 7,180 current input tokens/day versus a 12,340 baseline, a 41.8% reduction, with a target of at least 40%; these are README claims rather than an independent benchmark.

Link