QuantClaw
A plugin for the OpenClaw AI assistant that classifies each incoming request and routes it to a 4-bit, 8-bit or 16-bit precision model target, trading quality for latency and cost automatically instead of asking the user to choose.
QuantClaw evaluates OpenClaw workloads across 24 task types and several model families and sizes to decide which tasks tolerate lower-precision models without much quality loss, then applies that as a routing policy: a rule-based detector chain with a judge model as fallback picks a precision tier for each request in real time. The README's own benchmark shows high-precision-sensitive tasks, such as coding and safety-relevant work, keeping more of their score at higher precision, while lower-sensitivity tasks such as research or general comprehension tolerate quantization better.
Configuration is a JSON file naming the detector chain and the judge model's endpoint, and a bundled dashboard shows routing decisions, token counts, cost and live configuration changes as requests are served.
It is written in TypeScript and licensed under MIT.