UsageSniffer

Scan your local Claude Code, Codex, Opencode, Gemini CLI, Copilot CLI, Cursor, Aider, and Continue sessions and see visually what eats your tokens — skills, thinking, tools, context cache, plus dollar costs, anomalies, and cross-agent comparisons.

Zero-dependency Python. Your session logs never leave the machine.

Install

Requires Python 3.10+.

Windows (PowerShell):

winget install Python.Python.3.12
pip install usagesniffer        # or: pipx install usagesniffer
usagesniffer scan --top 10

macOS:

brew install python pipx
pipx install usagesniffer
usagesniffer scan --top 10

Linux:

pipx install usagesniffer        # or: pip install usagesniffer
usagesniffer scan --top 10

Or run from a checkout with no install: python3 -m usagesniffer scan --top 10.

OS notes

• Windows: Claude (%USERPROFILE%\.claude), Codex (%USERPROFILE%\.codex), Opencode (%USERPROFILE%\.local\share\opencode\opencode.db), Gemini, Copilot, and Continue resolve automatically. Cursor is read from %APPDATA%\Cursor\User. If your stores live elsewhere: usagesniffer scan --opencode-db PATH --cursor-dir DIR --aider-path FILE (Opencode also honors the OPENCODE_DB env var).
• macOS: Cursor is read from ~/Library/Application Support/Cursor/User; everything else is home-relative and just works.
• Linux: default paths as listed in "What it reads" below.

Quick start

usagesniffer scan --top 10                     # everything
usagesniffer scan --agents claude --top 5
usagesniffer session 05062c81                  # drill into one session (prefix match)
usagesniffer cost --top 10                     # dollar rollup + cache savings
usagesniffer cost --budget 50                  # exit 2 + alert when over $50
usagesniffer report -o report.html
usagesniffer watch --interval 30               # live burn (Ctrl-C to stop)
usagesniffer anomalies                         # outlier sessions vs agent baseline
usagesniffer skills-roi                        # per-skill cost table
usagesniffer compare                           # cross-agent efficiency
usagesniffer doctor                            # proactive waste audit (read-only, with fixes)
# filters work on most commands:
usagesniffer scan --since 2026-09-01 --project Sunrise --model sonnet

What it reads

Attribution buckets

Real totals always come straight from the logs. The buckets are token-weighted estimates (labelled as such in output):

• thinking — Claude thinking blocks / Codex reasoning_output_tokens / Opencode reasoning parts
• skills — SKILL.md loads, skill-instruction blocks, Skill-tool calls
• tools — tool_use + tool_result / exec payload share
• context_cache / context_write — cache read / creation tokens
• assistant_text, system, user, other

The second chart ("action share EXCLUDING context cache") is the useful one: cache reads dominate raw totals but are cheap; the action chart shows where your real budget goes.

Layout

usagesniffer/
  models.py            # SessionRecord / ScanResult
  analyze.py           # aggregation
  pricing.py           # per-model $/Mtok rates + cost rollups
  insights.py          # anomalies, skill ROI, cross-agent compare
  render.py            # terminal ASCII charts (stdlib only)
  report.py            # self-contained HTML report
  cli.py               # scan | top | session | cost | report | watch |
                       #   anomalies | skills-roi | compare
  parsers/
    claude.py codex.py opencode.py gemini.py
    copilot.py cursor.py aider.py continue_.py
tests/
  smoke_fixtures.py    # synthetic sessions through the real pipeline

Roadmap

• per-model pricing + $ cost rollup (cost, --budget)
• --watch live mode
• static HTML report export (report)
• prompt-caching savings estimate (cache_read vs full-price replay)
• 8-agent coverage (Claude, Codex, Opencode, Gemini, Copilot, Cursor, Aider, Continue)
• anomalies, skill ROI, cross-agent compare
• proactive doctor audit (skill weight, output hogs, model fit, cache reuse, disk)
• PyPI publish (pipx install usagesniffer)
• Homebrew / AUR packages — templates in packaging.md