Claude Code Cache Efficiency Analyzer

Based on https://habr.com/ru/companies/bitrix/articles/1008320/

CLI tool that reads Claude Code JSONL session files, computes prompt cache efficiency metrics using the Bitrix24/Habr formula, and displays results as rich terminal tables.

Formula

C = S × [(1−h) × P_miss + h × P_hit] + D × P_miss + O × P_out

Where: S = cacheable tokens, h = hit rate, D = dynamic input tokens, O = output tokens.

Install & Run

Via uvx (no install needed)

uvx claude-cache-analyzer

Via pip

pip install claude-cache-analyzer
claude-cache

From source

git clone https://github.com/AgiMateIo/claude-cache-analyzer.git
cd claude-cache-analyzer
uv sync
uv run python cli.py

Sample Output

╭──────────────────────────────────────────────────╮
│    Claude Code · Cache Efficiency Report         │
│    Project: all projects  ·  3 sessions analysed │
╰──────────────────────────────────────────────────╯

                    Summary
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━┓
┃ Metric                             ┃   Value ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━┩
│ Total actual cost                  │ $2.0190 │
│ Cost without cache                 │ $6.9472 │
│ Total savings                      │ $4.9283 │
│ Net savings (after write overhead) │ $4.7206 │
│ Savings %                          │   70.9% │
│ Avg cache hit rate                 │   85.5% │
│ Avg efficiency score               │    0.85 │
└────────────────────────────────────┴─────────┘

╭─────────────────────────── Tips ────────────────────────────╮
│ • Excellent hit rate (85.5%)! Cache is working efficiently. │
╰─────────────────────────────────────────────────────────────╯

Usage

Usage: claude-cache-analyzer [OPTIONS] [PROJECT_PATH]

 Analyze Claude Code prompt cache efficiency.

Arguments:
  [PROJECT_PATH]  Root of Claude data (~/.claude) or a specific project directory.

Options:
  -p, --project-name     TEXT     Filter by project directory name.
  -n, --top              INTEGER  Show only the N most recent sessions.
      --min-turns        INTEGER  Minimum number of turns to include a session. [default: 1]
  -s, --session          TEXT     Show detailed view for a specific session (full or partial ID).
  -g, --group-by-project          Group results by project.
      --export-json      PATH     Export raw metrics to a JSON file.
      --version                   Show version.
  -h, --help                      Show this message and exit.

Examples

# Analyze all sessions in ~/.claude
claude-cache-analyzer

# Last 5 sessions
claude-cache-analyzer --top 5

# Specific project
claude-cache-analyzer --project-name my-project

# Detailed view of a specific session (full or partial ID)
claude-cache-analyzer -s abcd1234

# Group by project
claude-cache-analyzer -g

# Export metrics to JSON
claude-cache-analyzer --export-json metrics.json

# Analyze a specific path
claude-cache-analyzer ~/.claude/projects/abc123

Metrics

Metric	Description
Cache hit rate	cache_read / (cache_creation + cache_read)
Actual cost	Real cost with cache pricing applied
Cost without cache	Hypothetical cost if all tokens were at input price
Savings	cost_no_cache - actual_cost
Net savings	Savings minus cache write overhead
Efficiency score	hit_rate × (cacheable / (input + cacheable)) — range [0..1]

Grades

Grade	Efficiency Score
A	≥ 0.70
B	≥ 0.50
C	≥ 0.30
D	≥ 0.10
F	< 0.10

Tests

uv run pytest -v