English | 简体中文

cc-speed

Analyze Claude Code token output speed from local JSONL logs.

What it does

cc-speed reads your local Claude Code conversation logs (~/.claude/projects/**/*.jsonl) and calculates token output speed (tokens/sec) by measuring the time between user and assistant messages.

It provides:

• Per-model statistics (median, avg, p10, p90)
• Daily trend sparklines in terminal
• Interactive HTML charts (scatter plot + histogram) with dark theme

Install

npm install -g cc-speed

Or run directly:

npx cc-speed

Usage

# Default: terminal table for the last 7 days
cc-speed

# Custom time range
cc-speed --days 30

# Generate HTML chart and open in browser
cc-speed --chart

# JSON output (for piping)
cc-speed --json

Options

Flag	Description	Default
-d, --days <n>	Number of days to analyze	7
-c, --chart	Generate HTML chart and open in browser	-
-j, --json	Output as JSON	-

Claude Code Slash Command

You can use cc-speed as a /speed command inside Claude Code:

# One-line install
bash <(curl -fsSL https://raw.githubusercontent.com/bryant24hao/cc-speed/main/install-skill.sh)

Or manually create ~/.claude/skills/cc-speed/SKILL.md:

---
name: speed
description: Analyze Claude Code token output speed
---

Run `npx cc-speed --chart` to analyze token output speed and generate charts.

Then type /speed in Claude Code to trigger it.

How it works

1. Scans ~/.claude/projects/**/*.jsonl for recent conversation logs
2. Pairs each assistant message with its preceding user message
3. Calculates output_tokens / (assistant_timestamp - user_timestamp)
4. Filters: output_tokens > 10, duration between 0.5s and 300s
5. Groups by model, computes statistics and trends

Note: The speed is an estimate based on timestamp deltas. It includes network latency and any processing overhead, so actual generation speed may be higher.

License

MIT