Less cloud context
Let local tools read first and return relevant excerpts. Avoid paying to send unnecessary source text to the cloud.
OPEN SOURCE · WINDOWS · MCP
Reduce paid cloud context with local search and analysis. Find relevant code in an index, inspect files on your machine, and send concise findings to your agent.
MIT licensed. Windows · MCP · CLI. Savings and latency depend on the task, index and hardware.
LOCAL AILet local tools read first and return relevant excerpts. Avoid paying to send unnecessary source text to the cloud.
Search an existing index by meaning or symbol, with source locations for the next step.
Usage notices distinguish estimated savings, additional overhead and unknown results.
INSIDE THE TOOLKIT
Three views of one system. Select a stage to see its purpose and boundary.
Conceptual stages, not live status. The signing view includes a planned SignPath step.
Read the full architecture →UNDERSTAND THE ESTIMATE
Compare against the context necessary for the request, not the combined size of every file opened locally.
Read the methodology →Illustrative only. All calculations stay in this page. No model call, analytics or account data.
THE HANDBOOK
Start with a supported workflow, then go deeper.
Requirements, installer choices and the first repository.
↗ 02CLI commands, index status and practical checks.
↗ 03What stays local, what clients receive, and release verification.
↗ 04Project boundaries, governance, support and contribution guides.
↗FAQ
No. The calling client receives results and selected excerpts; a cloud agent may forward them to its provider. See the privacy policy for the exact boundary.
No. If the source is already in context, or the request and result cost more than the necessary alternative, delegation may add overhead. Missing evidence means an unknown estimate.
Do not assume that. Current distribution uses an ECDSA-signed manifest. SignPath onboarding and a verified CI signing path must be completed before claiming an Authenticode-signed release.
Yes. They share the broker and model queue. Check status and coordinate long operations; do not stop the shared runtime or unload another session's model to troubleshoot your own project.