Hewn

Hewn

[](https://github.com/tommy29tmar/hewn/actions/workflows/ci.yml) [](LICENSE) [](https://github.com/tommy29tmar/hewn/releases) [](https://claude.com/claude-code)

**Make Claude Code go longer before limits hit.**

*Less output burn. Less drift. Faster long sessions.*

Claude Code users tend to stay in one long session — it's the practical path: no lost context, no re-loading files, no re-doing setup. But that's also where the cost hits hardest: repeated context, long answers, and output burn exactly when the budget starts to matter.

Hewn is built for that failure mode. It keeps Claude Code tighter over long sessions by routing each turn into a compact answer shape and then re-asserting that route every turn, so the model spends less output on restating context and more on the actual work.

Caveman already showed that compressed prompting helps. Hewn is tuned for a different bottleneck: long-session efficiency when Claude Code limits get tight.

Simple rule: if you want the most aggressive compression on one giant prompt, use Caveman. If you spend hours inside one Claude Code session and the limits are the pain, use Hewn.

No proxy. No telemetry. Default `claude` untouched.


This is the proof asset: the multi-turn benchmark where Hewn is strongest.

5-turn workflow, Claude Opus 4.7, same model and prompts across all four arms. Concept retention measured by transcript-aware LLM-as-judge.

Full methodology, raw snapshots, and