context-guru wants to stop paying for tools your agent never uses

A coding agent can carry tool definitions it never calls. context-guru 0.6.6 makes some of that baggage searchable on demand. The community project’s November 3 release introduces an opt-in ultrahigh preset that defers tools unused across the last three sessions, while adjusting prompt-cache keep-alive behavior to the kind of wait a thread is in.

context-guru is a proxy with client plugins, not a model. It sits between a coding harness and the upstream provider. In the new preset, eligible tools are sent with deferred loading on the Anthropic route, and the model uses a search tool to load a definition when needed. An MCP server is deferred as one unit; core operations such as Bash, Read, Edit and the project’s own tools remain available.

Experimental means watching the trade-off

The author reports that a live check reduced the cached prefix from 30,693 to 12,664 tokens with 12 built-in tools deferred. That is an attributed example, not a promised saving for another developer’s workload.

The release explicitly calls ultrahigh experimental and not recommended yet. In its reported tool-selection evaluation, Opus chose the right tool in 70 of 80 runs with deferral, versus 73 of 80 without. The notes also say the displayed net savings are too high because they omit a subsequent cache write, and that some searches can be counted twice. AITrending has not independently tested the feature or its accounting.

Deferral in this release is limited to Claude models through the Anthropic route. For GPT, GLM and Kimi, ultrahigh behaves like xhigh; installing the Codex bundle does not make this new deferral path apply to OpenAI models.

Cache pings follow the wait

For Anthropic’s five-minute cache, a thread waiting on its own tool can now receive up to eight keep-alive pings, while a thread that never called a tool gets none. Main-agent waits for background work or a person have different counts. These pings spend the user’s quota; the release says the new per-wait plan has not yet been measured on real traffic.

The project supplies statically linked Linux and macOS binaries, plus Claude Code plugin 0.6.6 and experimental Codex plugin 0.2.10. Its documented setup routes model traffic through a local proxy using the user’s upstream authentication; it does not provide free inference. Claude Code supports project routing, while Codex setup changes the user’s default provider and takes effect in a new session. Users need to understand that routing permission as well as the preset before trying the experiment.