How it works

Tokenistt sits between your editor and the Anthropic API as a transparent MCP proxy. Every prompt passes through it before being forwarded — adding zero latency to your perceived response time.

What gets analyzed

• –Prompt token count (input + system + tools)
• –Output token estimate and actual delta
• –Cache hit/miss status and prefix fingerprint
• –Model selection efficiency (over-provisioning detection)
• –Prompt structure waste (redundant instructions, repeated context)
• –Per-workspace cost attribution in real time

What never leaves your machine

Tokenistt does not log prompt content by default. Only metadata — token counts, model IDs, latency, cost — is sent to the analytics backend. Enable content_trace: true explicitly to capture full prompts for debugging.