pi Cost Tracker

A pi extension that tracks LLM API costs across all sessions. Every API call is logged to a local append-only ledger so you always know what you're spending.

Features

• Automatic tracking — every LLM call is recorded with token counts, cost breakdown, provider, and model
• Multi-client safe — uses append-only JSONL files, safe for concurrent writes from multiple pi instances
• /cost command — view today's cost or a multi-day range with breakdowns by model and provider
• Zero configuration — install and go, no API keys or setup required

Installation

From npm (recommended)

pi install npm:@ctogg/pi-cost-counter

From git

pi install git:github.com/cristeahub/pi-cost-counter

Local

pi install /path/to/pi-cost-counter

Quick test (temporary, current session only)

pi -e npm:@ctogg/pi-cost-counter
pi -e /path/to/pi-cost-counter

After installation, restart pi or run /reload to activate.

Usage

/cost Command

View cost summaries directly in pi:

/cost            Today's total + model breakdown
/cost 7d         Last 7 days
/cost 30d        Last 30 days
/cost 365d       Last year

Example output:

Cost for last 7 days (2026-04-11 → 2026-04-17)

  Total: $4.82  156.3K tokens · 87 calls

  Daily breakdown
  ────────────────────────────────────────────────────
  2026-04-11  $0.45
  2026-04-12  $1.23
  2026-04-13  —
  2026-04-14  $0.67
  2026-04-15  $0.89
  2026-04-16  $0.34
  2026-04-17  $1.24

  By model
  ────────────────────────────────────────────────────
  anthropic/claude-sonnet-4-5                    $3.41  120.5K tok · 62 calls
  anthropic/claude-haiku-3-5                     $0.89   28.1K tok · 18 calls
  google/gemini-2.5-pro                          $0.52    7.7K tok · 7 calls

Data Storage

Cost records are stored as JSONL (JSON Lines) files, organized by date:

~/.pi/cost-tracker/
  2026/
    04/
      17.jsonl
      18.jsonl
    05/
      01.jsonl

Record Format

Each line in a day file is a single JSON object representing one LLM API call:

{
  "ts": 1713369600000,
  "provider": "anthropic",
  "model": "claude-sonnet-4-5",
  "tokens": {
    "input": 1200,
    "output": 340,
    "cacheRead": 800,
    "cacheWrite": 0
  },
  "cost": {
    "input": 0.0036,
    "output": 0.0051,
    "cacheRead": 0.0008,
    "cacheWrite": 0,
    "total": 0.0095
  }
}

Field	Description
ts	Unix timestamp in milliseconds
provider	API provider (e.g. anthropic, google, openai)
model	Model identifier (e.g. claude-sonnet-4-5)
tokens.input	Input/prompt tokens
tokens.output	Output/completion tokens
tokens.cacheRead	Tokens served from prompt cache
tokens.cacheWrite	Tokens written to prompt cache
cost.input	Cost for input tokens (USD)
cost.output	Cost for output tokens (USD)
cost.cacheRead	Cost for cache-read tokens (USD)
cost.cacheWrite	Cost for cache-write tokens (USD)
cost.total	Total cost for this call (USD)

Querying Data Outside pi

Since the data is plain JSONL, you can query it with standard tools:

# Today's total cost
cat ~/.pi/cost-tracker/2026/04/17.jsonl | jq -s 'map(.cost.total) | add'

# Top models this month
cat ~/.pi/cost-tracker/2026/04/*.jsonl | jq -s 'group_by(.model) | map({model: .[0].model, total: map(.cost.total) | add}) | sort_by(-.total)'

# Total spend for April
cat ~/.pi/cost-tracker/2026/04/*.jsonl | jq -s 'map(.cost.total) | add'

# Number of API calls per day this month
for f in ~/.pi/cost-tracker/2026/04/*.jsonl; do
  echo "$(basename $f .jsonl): $(wc -l < $f) calls"
done

# Export to CSV
cat ~/.pi/cost-tracker/2026/04/*.jsonl | jq -r '[.ts, .provider, .model, .cost.total] | @csv'

How It Works

The extension hooks into pi's event system:

1. message_end — fires after every message (user, assistant, tool result). The handler filters for assistant messages, extracts the usage object that pi attaches to every LLM response, and appends a JSONL record to the appropriate day file.

2. /cost command — reads JSONL files for the requested date range, aggregates by day and by model, and displays a formatted summary.

Limitations

• Cost accuracy depends on pi's cost calculation — the extension records whatever pi reports in the usage.cost fields. If a model's pricing is not configured in pi, costs may show as zero even though tokens were consumed.
• Local timezone — day boundaries are based on your system's local time.
• No real-time sync — the /cost command reads files at invocation time. It does not live-update while other pi clients are running. Run it again to see the latest.

License

MIT