Docs
GitHubStar

Stop hitting your token limits.

Session burn checks to save you tokens.
Tiny, fast, free, local. Open source.

curl -fsSL https://antiburn.ai/install.sh | sh
The antiburn popover: live Claude limit meters above today's coding-agent sessions, each showing its cost and how many burn checks it passed.
  • Stay on the good side of your reset line.
  • Tweak your config so it works for you not the model providers.
  • See the habits that are costing you, and how to improve them.
The context tab: a chart of a real session's context climbing in steps to 218k tokens, dropping at a compaction, then climbing again, with token traffic and sub-agent bursts underneath, and a key reading 218.4k context, 498.9k in, 103.9k out, 1 compaction.The cost tab: spend split by parent agent and sub-agents, then by input, output, cache read and cache write, with the session's cost per million tokens.The tools tab: tokens burned by items loaded but never called, then every skill, MCP and tool loaded into the session with its token cost and whether it was used.

$419.29/month

Sessions running into the expensive and dumb region

LLMs have to re-send every token in the context window for every request, and caching only reduces that impact slightly. The also get less intelligent with more tokens to consider. Compaction works these days, and the models are best below about 300k tokens. Change your auto-compact settings.

Sessions past 400k context
72
Deepest session
998k tokens
Spend in deep turns
71%
Monthly impact (API pricing equiv)
$419.29
  • antiburn checks your sessions to see how deep they go, and estimates the cost of that.
  • Set "env.CLAUDE_CODE_AUTO_COMPACT_WINDOW": "400000" in ~/.claude/settings.json.
  • Also build a habit of compacting manually using /compact.

$419.29/month

Sessions running into the expensive and dumb region

LLMs have to re-send every token in the context window for every request, and caching only reduces that impact slightly. The also get less intelligent with more tokens to consider. Compaction works these days, and the models are best below about 300k tokens. Change your auto-compact settings.

Sessions past 400k context
72
Deepest session
998k tokens
Spend in deep turns
71%
Monthly impact (API pricing equiv)
$419.29
  • antiburn checks your sessions to see how deep they go, and estimates the cost of that.
  • Set "env.CLAUDE_CODE_AUTO_COMPACT_WINDOW": "400000" in ~/.claude/settings.json.
  • Also build a habit of compacting manually using /compact.

  1. antiburn runs light

    We hate heavy apps, so we keep it below 300 MB memory and 5% CPU.

  2. Keep your data local

    Everything runs on your device, no server or cloud needed.

  3. Track it over time

    See the burn impact of new models, new habits, new settings.

Used by dev teams at

BuildPassMindset HealthAP0SitemateEQLGreat QuestionFGPRome2riohipagesAccolade Pro