cc-flytrap

Claude Code optimizer. ~95% less tokens → faster + cheaper.

Install

curl -fsSL https://raw.githubusercontent.com/adiled/cc-flytrap/main/install.sh | bash

Use

After install, Claude uses it automatically. Restart Claude if already running.

Commands

Command	Description
ccft start	Start optimizer
ccft stop	Stop
ccft status	Check running
ccft install	Install as service
ccft update	Update
ccft brainrot	Vibe-check your sessions (see below)
ccft perf	Is ccft adding latency? (see below)

Brainrot

Two-axis vibe-check on your sessions. Splits bot drift (output collapse, latency choke) from driver drift (input bloat, rapid-fire, session sprawl). Tells you whose fault.

Behavioural only — derived from velocities, latencies, permutations, and volumetrics in the ledger. Never reads your content.

Command	What
ccft brainrot	Today's dashboard
ccft brainrot week	7-day rollup
ccft brainrot replay --follow	Live tail of new requests
ccft brainrot diff today yesterday	Compare two ranges
ccft brainrot session [sid]	Drill into one session
ccft brainrot score	One-liner for status bars

Perf

ccft's own observability layer. Answers: is ccft slowing my requests down?

ccft perf · today
─────────────────
wall        p50  1.2s   p95  8.4s   p99  31s
upstream    p50  760ms  p95  7.1s   p99  26s
pre         p50  480ms  p95  1.2s   p99  5.4s
ccft        p50  2ms    p95  8ms    p99  14ms

verdict     ccft contributes ~0.2% of wall time. not the bottleneck —
            slowness is upstream.

Bucket	Meaning
wall	total time the flow spent under ccft (request received → response complete)
upstream	streaming response duration (api.anthropic.com → us)
pre	wall − upstream — wait for first byte (mostly API TTFT)
ccft	measured ccft internal processing (json parse, modify, hooks)

The verdict gives a plain-language answer based on the median ccft time, both in absolute milliseconds and as a percentage of wall time.

Config

~/.config/ccft/ccft.json:

{
  "port": 7178,
  "host": "127.0.0.1",
  "system_override": "",
  "pain": false,
  "ledger": true
}

Key	Default	Meaning
pain	false	false = trim bloated system prompts. true = leave them alone (passive observer).
ledger	true	Record per-request telemetry to ~/.local/share/ccft/ledger.jsonl. Set false to disable.
system_override	""	Custom system prompt to inject. Empty = use built-in default. Injected regardless of pain.

Logs

~/.local/share/ccft/logs/