llm-usage-metrics

CLI to aggregate local LLM usage from:

• ~/.pi/agent/sessions/**/*.jsonl
• ~/.codex/sessions/**/*.jsonl
• OpenCode SQLite DB (auto-discovered or provided via --opencode-db)

Reports are available for daily, weekly (Monday-start), and monthly periods.

Project documentation is available in docs/.

Built-in adapters currently support 3 sources: .pi, .codex, and OpenCode SQLite. The codebase is structured to add more sources (for example Claude/Gemini exports) through the SourceAdapter pattern. See CONTRIBUTING.md.

Install

npm install -g llm-usage-metrics

Or run without global install:

npx --yes llm-usage-metrics daily

(npx llm-usage daily works when the project is already installed locally.)

Runtime notes:

• OpenCode parsing requires Node.js 24+ (node:sqlite).
• Bun is supported for dependency/scripts workflow, but OpenCode report runs should use Node-based CLI execution.
• Example local execution against built dist: node dist/index.js daily --source opencode --opencode-db /path/to/opencode.db

Update checks

When installed globally, the CLI performs a lightweight npm update check on startup.

Behavior:

• uses a local cache (<platform-cache-root>/llm-usage-metrics/update-check.json; defaults to ~/.cache/llm-usage-metrics/update-check.json on Linux when XDG_CACHE_HOME is unset) with a 1-hour default TTL
• optional session-scoped cache mode via LLM_USAGE_UPDATE_CACHE_SCOPE=session
• skips checks for --help / --version invocations
• skips checks when run through npx
• prompts for install + restart only in interactive TTY sessions
• prints a one-line notice in non-interactive sessions

To force-skip startup update checks:

LLM_USAGE_SKIP_UPDATE_CHECK=1 llm-usage daily

Runtime environment overrides

You can tune runtime behavior with environment variables:

• LLM_USAGE_SKIP_UPDATE_CHECK: skip startup update check when set to 1
• LLM_USAGE_UPDATE_CACHE_SCOPE: cache scope for update checks (global default, session to scope by terminal shell session)
• LLM_USAGE_UPDATE_CACHE_SESSION_KEY: optional custom session key when LLM_USAGE_UPDATE_CACHE_SCOPE=session (defaults to parent shell PID)
• LLM_USAGE_UPDATE_CACHE_TTL_MS: update-check cache TTL in milliseconds (clamped: 0..2592000000; use 0 to check on every CLI run)
• LLM_USAGE_UPDATE_FETCH_TIMEOUT_MS: update-check network timeout in milliseconds (clamped: 200..30000)
• LLM_USAGE_PRICING_CACHE_TTL_MS: pricing cache TTL in milliseconds (clamped: 60000..2592000000)
• LLM_USAGE_PRICING_FETCH_TIMEOUT_MS: pricing fetch timeout in milliseconds (clamped: 200..30000)
• LLM_USAGE_PARSE_MAX_PARALLEL: max concurrent file parses per source adapter (clamped: 1..64)

Example:

LLM_USAGE_PARSE_MAX_PARALLEL=16 LLM_USAGE_PRICING_FETCH_TIMEOUT_MS=8000 llm-usage monthly

Usage

Daily report (default terminal table)

llm-usage daily

Weekly report with custom timezone

llm-usage weekly --timezone Europe/Paris

Monthly report with date range

llm-usage monthly --since 2026-01-01 --until 2026-01-31

Markdown output

llm-usage daily --markdown

JSON output

llm-usage daily --json

Offline pricing (use cached LiteLLM pricing only)

llm-usage monthly --pricing-offline

Override pricing URL

llm-usage monthly --pricing-url https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json

Pricing behavior notes:

• LiteLLM is the active pricing source.
• explicit costUsd: 0 events are re-priced from LiteLLM when model pricing is available.
• when pricing cannot be loaded from LiteLLM (or cache in offline mode), report generation fails fast.

Custom session directories

llm-usage daily --pi-dir /path/to/pi/sessions --codex-dir /path/to/codex/sessions

Or use generic source-id mapping (repeatable):

llm-usage daily --source-dir pi=/path/to/pi/sessions --source-dir codex=/path/to/codex/sessions

Directory override rules:

• --source-dir is directory-only (currently pi and codex).
• --source-dir opencode=... is invalid and points to --opencode-db.
• --opencode-db <path> sets an explicit OpenCode SQLite DB path.

OpenCode DB override:

llm-usage daily --opencode-db /path/to/opencode.db

OpenCode path precedence:

1. explicit --opencode-db
2. deterministic OS-specific default path candidates

Backfill example from a historical DB snapshot:

llm-usage monthly --source opencode --opencode-db /archives/opencode-2026-01.db --since 2026-01-01 --until 2026-01-31

OpenCode safety notes:

• OpenCode DB is opened in read-only mode
• unreadable/missing explicit paths fail fast with actionable errors
• OpenCode CLI is optional for troubleshooting and not required for runtime parsing

Filter by source (--source)

Use --source to limit reports to one or more source ids.

Supported source ids:

• pi
• codex
• opencode

Behavior:

• repeatable or comma-separated (--source pi --source codex or --source pi,codex)
• case-insensitive source id matching
• unknown ids fail fast with a validation error

Examples:

# only codex data
llm-usage monthly --source codex

# only pi data
llm-usage monthly --source pi

# only OpenCode data
llm-usage monthly --source opencode

# multiple sources
llm-usage monthly --source pi --source codex
llm-usage monthly --source pi,codex

# OpenCode source with explicit DB path
llm-usage monthly --source opencode --opencode-db /path/to/opencode.db

Filter by provider (--provider)

Use --provider to keep only events whose provider contains the filter text.

Behavior:

• case-insensitive substring match
• optional flag (when omitted, all providers are included)
• works together with --source and --model

Examples:

# all OpenAI-family providers
llm-usage monthly --provider openai

# GitHub Models providers
llm-usage monthly --provider github

# source + provider together
llm-usage monthly --source codex --provider openai

Filter by model (--model)

--model supports repeatable and comma-separated filters. Matching is case-insensitive.

Per filter value:

• if an exact model id exists in the currently selected event set (after source/provider/date filtering), exact matching is used
• otherwise, substring matching is used

Examples:

# substring match (all Claude-family models)
llm-usage monthly --model claude

# exact match when present
llm-usage monthly --model claude-sonnet-4.5

# multiple model filters
llm-usage monthly --model claude --model gpt-5
llm-usage monthly --model claude,gpt-5

# source + provider + model together
llm-usage monthly --source opencode --provider openai --model gpt-4.1

Per-model columns (opt-in detailed table layout)

Default output is compact (model names only in the Models column).

Use --per-model-columns to render per-model multiline metrics in each numeric column:

llm-usage monthly --per-model-columns
llm-usage monthly --markdown --per-model-columns

Output features

Terminal UI

The CLI provides an enhanced terminal output with:

• Boxed report header showing the report type and timezone
• Session summary displayed at startup (session files and event counts per source)
• Pricing source info indicating whether data was loaded from cache or fetched remotely
• Environment variable overrides displayed when active
• Models displayed as bullet points for better readability
• Rounded table borders and improved color scheme

Example output:

ℹ Found 12 session file(s) with 45 event(s)
•   pi: 8 file(s), 32 events
•   codex: 4 file(s), 13 events
ℹ Loaded pricing from cache

┌──────────────────────────────────────────────────────────┐
│ Monthly Token Usage Report (Timezone: Africa/Casablanca) │
└──────────────────────────────────────────────────────────┘

╭────────────┬──────────┬──────────────────────╮
│ Period     │ Source   │ Models               │
├────────────┼──────────┼──────────────────────┤
│ Feb 2026   │ pi       │ • gpt-5.2            │
│            │          │ • gpt-5.2-codex      │
╰────────────┴──────────┴──────────────────────╯

Report structure

Each report includes:

• source rows (pi, codex, opencode) for each period
• a per-period combined subtotal row (only when multiple sources exist in that period)
• a final grand total row across all periods

Columns:

• Period
• Source
• Models
• Input
• Output
• Reasoning
• Cache Read
• Cache Write
• Total
• Cost

Development

bun install
bun run lint
bun run typecheck
bun run test
bun run format:check