contextburn

[contextburn: real output over 24 hours — useful work 0.18% of tokens, context re-reading 98.4%, cost-weighted useful work 6.7%, one useful token costs 555 paid tokens]

[DOI 10.5281/zenodo.22712985]

contextburn reads the transcripts Claude Code already writes on your machine and tells you what share of the tokens you paid for became model output — and how much was the agent re-reading context it had already sent.

Token counters answer "how much did I spend?". This answers "how much of it was work?" — a normalised share, so it can be compared across sessions, models and ways of working.

Try it

cp bin/contextburn ~/bin/contextburn && chmod +x ~/bin/contextburn   # python3 only, no dependenciescontextburn detail 24

Demo

[contextburn detail 72 over the 36 experiment runs: 168 sessions, useful work 1.44% of tokens, context re-reading 94.4%, cost-weighted 28.8%]

Real output over the session logs of the 36 runs behind the U-curve report — nothing else on the machine. Video with DOI: 10.5281/zenodo.22713920. The runs themselves are open: Hugging Face (DOI 10.57967/hf/10366) · Kaggle · OSF (DOI 10.17605/OSF.IO/5QTWY).

Why two numbers

[Same 12 tasks, one long session versus twelve short, 3 runs each: token efficiency 1.11% vs 1.12%, no difference; cost-weighted efficiency 31.6% vs 24.6%, seven points apart]

• By tokens the share barely moves. Every agent step resends the accumulated context, so re-reading dominates whatever you do — it describes the agent.
• Cost-weighted the share does move, because cached reads are priced far below fresh input and output. It depends on how you run sessions — it describes you.

The comparison above comes from a controlled experiment with its dataset and analysis scripts: Clear Every Third Task: A Measured U-Curve in the Context Economy of Coding Agents.

How it counts

• Reads local Claude Code transcripts (~/.claude/projects/**/*.jsonl). Nothing leaves the machine — no network calls at all.
• Deduplicates usage records by message id and keeps the element-wise maximum. A streaming runtime writes an early snapshot and a final record for the same call: counting both double-counts it, keeping only the first halves the output.
• Weights the cost share with per-model prices kept at the top of bin/contextburn. Update them there when they change.

Commands

Configuration

The language file exists because the menu-bar app is launched from Finder, where environment variables never reach it: echo ru > ~/.config/contextburn/lang switches both the app and the CLI.

MCP server

Let the agent read its own run efficiency mid-session. The package ships a dependency-free MCP server (stdio) with two tools: run_efficiency returns the shares as structured data, and spend_breakdown returns the full report.

claude mcp add contextburn -- uvx contextburn mcp

Or install it as a Claude Code plugin, which registers the same server:

/plugin marketplace add arsentev-ai/contextburn/plugin install contextburn@contextburn

Editor extensions

• VS Code-compatible editors (VSCodium, Cursor, Windsurf, Gitpod…) — Open VSX: arsentev-ai.contextburn. A status bar meter over the local CLI; source in editors/vscode.
• Raycast — source in editors/raycast, Store submission pending.

Menu-bar app (macOS)

app/main.swift is a small status-bar app. It polls contextburn --json once a minute and shows the current burn rate with an hourly graph; click a bar to see that hour's breakdown.

swiftc -O -o ContextBurn app/main.swift

Set CONTEXTBURN_BIN=/path/to/contextburn if the CLI is not in ~/bin or the usual Homebrew paths.

Limits

• Claude Code transcripts only, for now.
• The cost-weighted share is only as current as the price table in bin/contextburn.

Citing

Software DOI (all versions): 10.5281/zenodo.22712985. GitHub's "Cite this repository" button gives the reference; metadata is in CITATION.cff.

Author

Evgenii Arsentev — arsentev.ai · ORCID 0000-0002-9120-7298

This project was published as tokmon on its first day and renamed to avoid confusion with unrelated tools of that name; TOKMON_* environment variables still work.

License

MIT — see LICENSE.