contextburn MCP server

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 dependencies
contextburn detail 24

Demo

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

re-reading dominates whatever you do — it describes the agent.

• By tokens the share barely moves. Every agent step resends the accumulated context, so

output. It depends on how you run sessions — it describes you.

• Cost-weighted the share does move, because cached reads are priced far below fresh input and

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

machine — no network calls at all.**

• Reads local Claude Code transcripts (~/.claude/projects//.jsonl). **Nothing leaves the

writes an early snapshot and a final record for the same call: counting both double-counts it, keeping only the first halves the output.

• Deduplicates usage records by message id and keeps the element-wise maximum. A streaming runtime

there when they change.

• Weights the cost share with per-model prices kept at the top of bin/contextburn. Update them

Commands

Configuration

Shortened. The full README is on GitHub.