contextburn
SAFEgrade B · trust 89/100How much of what your coding agent spends is actually work: the share of paid tokens that became output, not context re-reading. Local, no network calls.
claude codecoding agentscontext windowcost efficiencyllmobservabilitytoken usageOverview
From the repository's own README, as read at the audited commit.
<p align="center"> <img src="https://raw.githubusercontent.com/arsentev-ai/contextburn/main/assets/readme/hero.svg" width="100%" alt="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"></p><p align="center"><a href="https://doi.org/10.5281/zenodo.22712985"><img src="https://zenodo.org/badge/DOI/10.5281/zenodo.22712985.svg" alt="DOI 10.5281/zenodo.22712985"></a></p>**contextburn** reads the transcripts Claude Code already writes on your machine and tells you whatshare of the tokens you paid for became model output — and how much was the agent re-readingcontext it had already sent.Token counters answer *"how much did I spend?"*. This answers *"how much of it was work?"* — anormalised share, so it can be compared across sessions, models and ways of working.## Try it```bashcp bin/contextburn ~/bin/contextburn && chmod +x ~/bin/contextburn # python3 only, no dependenciescontextburn detail 24```## Demo<p align="center"> <img src="https://raw.githubusercontent.com/arsentev-ai/contextburn/main/assets/readme/demo.gif" width="100%" alt="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%"></p>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](https://doi.org/10.5281/zenodo.22713920). The runs themselves are open: [Hugging Face](https://huggingface.co/datasets/arsentev-ai/context-ucurve-coding-agents) (DOI 10.57967/hf/10366) · [Kaggle](https://www.kaggle.com/datasets/arsentevai/context-u-curve-of-coding-agents-36-runs) · [OSF](https://osf.io/5qtwy/) (DOI 10.17605/OSF.IO/5QTWY).## Why two numbers<p align="center"> <img src="https://raw.githubusercontent.com/arsentev-ai/contextburn/main/assets/readme/two-numbers.svg" width="100%" alt="Same 12 tasks, one long session versus twel
Connect
Built from this server's own package name, version and transport as found in its source — not copied from anyone's documentation, so it cannot drift against a page we do not control.
claude mcp add contextburn -- uvx contextburn==0.2.1 mcp
Details
- Source
- arsentev-ai/contextburn
- npm
contextburn@0.1.0- PyPI
contextburn- Transports
- stdio
- Tools
- Registered at runtime; the list could not be read from source.
- License
- MIT
- Stars
- 0 · pushed 0d ago
Trust audit
Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | PASS |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | PASS |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- none-observed
- Network
- none-observed
- Shell
- none-observed
- Dependencies
- not all pinned
- Secrets in source
- none-found
Findings (5)
.vscodeignore
@raycast/api, @raycast/eslint-config, eslint, prettier, typescript
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](https://doi.org/10.5281/zenodo.22713920). The runs t
and cost followed a U-curve in session length — report: DOI [10.5281/zenodo.22699668](https://doi.org/10.5281/zenodo.22699668).
Gates applied: no_behavioural_pass.
Audited 2026-09-14 · audit v0.4.0 · source sha aaacd93da484 · full audit: observations/trust-audit/mcp-server/arsentev-ai__contextburn.json · Report an issue or request a re-scan
Audit history
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-09-14 | aaacd93da484 | SAFE | B | 89 | source changed, verdict held |
| 2026-09-12 | 1f13b187797a | SAFE | B | 89 | first audit |
Alternatives
Other servers in the same categories, safer ones first.
Questions
What is the contextburn MCP server?
How much of what your coding agent spends is actually work: the share of paid tokens that became output, not context re-reading. Local, no network calls.
Is contextburn safe to connect to an agent?
The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean server reads B.
What credentials does contextburn need?
No credential environment variables were found in its source, so it appears to need none.
How does contextburn run?
It speaks stdio, so it runs as a local process your client starts. It is published on npm as contextburn at 0.1.0.
How current is this page?
The grade is for one exact copy of the source (aaacd93da484), read on 2026-09-14. The repository is watched and re-audited when it changes.
Provenance: OBSERVED · read 2026-09-14 · job trust-audit-2026-09-14