Cache hits, cache misses, cache writes, saved USD.

Watch one thing deeply: LLM prompt-cache efficiency.

`catop` is a compact terminal monitor for teams running coding agents, RAG loops, eval runners, and LiteLLM Proxy traffic.

View on GitHub Quick start

The first screen is the product

Every number answers a cache question.

read cache hit tokens

Input tokens served from provider-side cache.

write cache creation tokens

Input tokens that created or refreshed cache entries.

miss normal input billing

Tokens that did not benefit from cache pricing.

saved estimated USD saved

Calculated from the LiteLLM model price map.

Install, point, watch

Run from source or from GitHub.

python -m pip install "catop @ git+https://github.com/Harzva/catop-cachehit.git"
catop --demo --once
catop --scan-agents --window today
catop --litellm-proxy-log --group-by session

Scope stays narrow.

`catop` is not a tracing platform. It is a cache-hit monitor that keeps read/write/miss/savings visible while agents are running.

Real data sources

LiteLLM Proxy, Claude Code, Codex, and JSONL.

LiteLLM Proxy callback

Writes successful request usage to JSONL without storing prompts.

Local coding agents

Scans Claude Code and Codex session stores when token usage is present.

Provider pricing

Uses LiteLLM's price map with a local TTL cache and offline fallbacks.

Screenshots

Dense terminal views, not a dashboard maze.

catop grouped cache dashboard — Grouped by agent, provider, model, and project.

catop single session detail — Single session detail with ordered cache events.

catop LiteLLM Proxy callback flow — LiteLLM Proxy callback writes cache metadata to JSONL.