claudeboard

A live web dashboard for Claude Code sessions. Watches ~/.claude/projects/ and tells you what every terminal tab is doing, what each session has cost, and how your month-to-date burn compares to the $200 Max plan.

Run

uv sync
uv run claudeboard

Open http://localhost:8765.

For one-click Sonnet session summaries, set ANTHROPIC_API_KEY in the environment.

What it shows

• Sidebar of every session with busy / idle / dead status, derived from ps+lsof (which claude processes are pointing at which cwd) with mtime fallback.
• Per-session cost at real Anthropic list prices for the actual model used. Header shows month-to-date vs the $200 Max plan.
• Per-session detail: token + cost mix (in / out / cache_r / cache_w), git diff and commits since session start, files edited/written/read with op counts, filtered transcript.
• Click the ↗ on any live row to bring its iTerm2 or Terminal.app tab to the foreground (macOS, AppleScript).
• Click SPEND in the header for a comprehensive view: presets (1h / 24h / 7d / 30d / MTD) or a custom datetime range, stacked time-series chart by model, breakdowns by token type / model / project, and the top sessions in the window.

Pricing notes

The four token types in one paragraph: every Claude Code turn re-sends the whole conversation + system prompt + tool definitions, so volume is dominated by cache reads at $1.50 / MTok (Opus). The bill is dominated by output at $75 / MTok (50× the cache-read rate, even though output is < 1% of volume) and cache writes at $18.75 / MTok (the 25%-premium-on-input cost of creating the cache so reads can be cheap).

sessions.py carries the PRICING table; unknown models fall back to Sonnet rates. Update when Anthropic changes prices.

Develop

uv sync --extra dev
uv run pre-commit install
uv run pytest
uv run ruff check
uv run ruff format

Runtime is stdlib only — no third-party Python deps. CI runs ruff + pytest on Python 3.10 / 3.11 / 3.12.