savings

Token Saver

boltToken Saver

Compress tool output (RTK)

git/grep/ls/tree/logs → 60-90% fewer input tokens

Compress context (Headroom)

Checking…

Compress prompts via /v1/compress before routing to the model

Compress LLM output (Caveman)

Terse-style system prompt → ~65% fewer output tokens (up to 87%)

Lazy senior dev (Ponytail)

Bias the model toward minimal code: YAGNI, reuse stdlib, deletion over addition

Custom system prompt

Inject your own text into the system message of every request

Tool relevance ranking

Drop tool definitions that share no word with your recent messages. Only affects requests with many tools; keeps at least half.

Filter MCP tool schemas

Static config-driven exclusion of irrelevant MCP server schemas

BM25 tool relevance

Per-turn BM25 ranking — sends only the most relevant schemas (zero new deps)