Releases·Tool Release
Claude Code transcript audit finds 4 token leaks; fixes cut cost per call 40%
A developer analyzed Claude Code's per-call token logs and found most cost came from context bloat, mis-inherited model assignments, and repeated file reads — not output verbosity. After four fixes, cost per call fell 40% and subagent spend dropped 77%. The audit and fixes are packaged as a local-only plugin, /scrooge.
9 engagement1 sourceSun, Oct 4, 2026, 05:10 AM
AI Overview
- Audit of ~/.claude/projects/*.l found 966k-token context re-read every call, driving 27% of all cache reads.
- 2 of 3 subagent runs defaulted to Opus because no model: was set in agent frontmatter.
- 9 of 10 sessions started one folder above the repo, so CLAUDE.md, agent configs, and pinned Sonnet never loaded.
- A 145 KB source file was read in full 44 times; visible assistant text was under 5% of output tokens.
- Fixes via CLAUDE_CODE_AUTO_COMPACT_WINDOW=300000, model pinning, in-repo sessions, and a PreToolUse Read hook cut subagent spend 77%.
- Author packaged the audit as a local-only Python plugin, /scrooge (github.com/mdmudassirahmed/scrooge).
Related
Releases·Tool Release
Developer builds open-source profiler for Claude Code, finds $15k in wasted tokens from re-reads and failed commands
A developer analyzed 60 sessions of Claude Code transcripts and built a free open-source profiler to identify token waste. Key findings: a 5,649-line progress doc was read 118 times, and 455 failed tool calls cost ~$201 in retry billing. Caching saved an estimated $15k in API costs.
A developer analyzed 60 sessions of Claude Code transcripts and built a free open-source profiler to identify token waste. Key findings: a 5,649-line progress doc was read 118 times, and 455 failed tool calls cost ~$201 in retry billing. Caching saved an estimated $15k in API costs.
21029 engagement3 sourcesredditThu, Jul 23, 2026, 10:04 AM
Social·Community
56% of my Claude Code usage was Claude re-reading the conversation
A user priced six months of personal Claude Code transcripts at API rates and found only 20% of spend went to useful work, while 56% covered re-reading the conversation, since every call resends full context. The post offers mitigation settings and promises follow-up before/after numbers.
A user priced six months of personal Claude Code transcripts at API rates and found only 20% of spend went to useful work, while 56% covered re-reading the conversation, since every call resends full context. The post offers mitigation settings and promises follow-up before/after numbers.
230 engagement1 sourceredditTue, Oct 6, 2026, 11:17 PM
Social·Community
Developer tests 3 Claude Code plugins to cut multi-agent token costs
A developer building an internal CRM with a multi-agent Claude Code setup tested three plugins aimed at reducing token costs. Analysis showed cache reads, not generated code, drove most token spend, since each turn resends growing context across agents. The post shares which cost-reduction approaches actually worked for practitioners running similar agent loops.
A developer building an internal CRM with a multi-agent Claude Code setup tested three plugins aimed at reducing token costs. Analysis showed cache reads, not generated code, drove most token spend, since each turn resends growing context across agents. The post shares which cost-reduction approaches actually worked for practitioners running similar agent loops.
27 engagement1 sourceredditMon, Aug 31, 2026, 03:57 PM
Products·launched
jev-gate: Claude Code plugin that routes tasks across Haiku/Sonnet/Opus via a cheap decision model
jev-gate is a Claude Code plugin where a cheap decision model decides whether each task needs Haiku, Sonnet, or Opus, while Sonnet coordinates. It also includes an extractive compactor that replaces auto-compact with a digest plus recent tail without calling a summarizer. It targets Claude Code users wanting to cut token costs, though no savings are proven yet.
jev-gate is a Claude Code plugin where a cheap decision model decides whether each task needs Haiku, Sonnet, or Opus, while Sonnet coordinates. It also includes an extractive compactor that replaces auto-compact with a digest plus recent tail without calling a summarizer. It targets Claude Code users wanting to cut token costs, though no savings are proven yet.
2 engagement1 sourceredditTue, Sep 29, 2026, 04:08 AM
Releases·Tool Release
We captured raw Claude Code API traffic: where tokens and budget actually go (breakdown of 5 request-level components)
Developers captured raw Claude Code API traffic via local proxy to analyze where tokens and cost go in multi-turn sessions. They identified five cost drivers, including MCP tool schemas consuming 15k-25k input tokens per turn and prompt cache misses that cut the 90% discount. They released an open-source tool, cost-xray, that attributes tokens and costs to individual sources for Claude Code and Codex.
Developers captured raw Claude Code API traffic via local proxy to analyze where tokens and cost go in multi-turn sessions. They identified five cost drivers, including MCP tool schemas consuming 15k-25k input tokens per turn and prompt cache misses that cut the 90% discount. They released an open-source tool, cost-xray, that attributes tokens and costs to individual sources for Claude Code and Codex.
2 engagement1 sourceredditFri, Sep 11, 2026, 03:57 PM