New releasev0.13.16Sep 11, 2026

New version of Boost releasedCopilot observability and Windows sync fixes

Keep the work. Cut the context bill.

Boost saves roughly 13% of a coding-agent sessionby removing repetitive tool output before it enters the model context.

~13%

less session cost

Typical saving; actual results vary by workflow.

No workflow change

Keep your agent, model, terminal, and commands.

Nothing is lost

Every compression is local, measured, and reversible.

Workflow proof

See the saving, task by task.

Each demo runs the same task with and without Boost, then publishes the complete token, time, and cost result.

Codebase exploration

Asking a question costs 50% less with Boost.

An agent traced Minikube’s missing-container-runtime preflight flow. BoostGraph replaced the long grep-and-read loop with structural code search, returning the same answer with fewer calls and less context.

Saving
51% less
Repository
minikube
Harness
Claude Code
Measured Minikube question results with and without BoostGraph
MetricWithout BoostWith BoostImprovement
Cost$0.2189$0.107151% lower
Time1m 58s37s3.2× faster
Input tokens2.3K63273% fewer
Output tokens10.7K3.3K69% fewer
Cache read689.2K180.5K74% fewer
Cache write68.8K36K48% fewer
File upload

Uploading a file costs 52% less with Boost.

An agent attached a large source file to work through a change. Boost compacted the surrounding tool output so the upload and follow-up reads carried the same workflow with less context.

Saving
52% less
Harness
Claude Code

Next in the lab

More real workflows, no mock data.

Videos and full telemetry in progress

Developing a feature

Benchmark pending

Fixing a bug

Benchmark pending

Debugging CI

Benchmark pending