How much have you actually spent on Claude Code? Which projects burn the most tokens? Which tools is Claude calling for you most often? Which single prompt cost you $1.50?

This is a small Rill project that turns the local Claude Code session logs on your laptop into interactive dashboards. No data leaves your machine — Rill reads the JSONL transcripts directly out of ~/.claude/projects/.

Two dashboards on the same data:

Claude Code Usage Overview (canvas) — an at-a-glance report

• KPI cards: total cost, total messages, tokens (output / cache read), unique sessions, unique projects, avg cost per session
• Tokens and messages over time, broken down by model
• Top projects, top sessions, and tools ranked by cost
• Donut of messages by role, stacked bars of messages by model
• A per-message detail table so you can read every prompt and assistant reply with its cost

Claude Code Cost & Usage Explorer (explore) — slice and drill

• Filter by project, model, role, entrypoint (cli / vscode), permission mode, tool name, sub-agent flag, and more
• Group by any dimension; the time series follows your filter
• Pivot to a raw-message table and inspect single API calls

You need Rill installed locally. If you don't have it:

curl https://rill.sh | sh

Then:

rill start https://github.com/rilldata/claude-usage.git

That's it. Open http://localhost:9009. The first time you run, Rill reads every JSONL file under ~/.claude/projects/ and materializes a DuckDB table — takes a few seconds. After that, navigate to Claude Code Usage Overview or Claude Code Cost & Usage Explorer in the sidebar.

Whenever you want fresher data, save any file in the project (or click "Refresh model" in the UI) and Rill reloads the JSONL files.

models/
  claude_messages.yaml         Reads ~/.claude/projects/*/*.jsonl, extracts every
                               user/assistant message with full text, tool calls,
                               thinking, tokens, and cost.
  claude_model_pricing.yaml    Per-model price table (input/output/cache rates).
                               Edit this if Anthropic changes prices.
metrics/
  claude_usage.yaml            ~25 dimensions and ~17 measures including
                               total_cost_usd, max_cost_per_message,
                               avg_cost_per_user_prompt.
dashboards/
  claude_usage_canvas.yaml     The overview canvas.
  claude_usage_explore.yaml    The drill-down explore.

• What did I actually spend? Total Cost (USD) KPI, defaults to all time.
• Which project was most expensive? Top Projects leaderboard, sorted by cost.
• What's my single most expensive turn ever? Sort the per-message table by cost descending.
• Which tool calls drive cost? "Cost by Tool Combination" leaderboard, or filter tool_names in the explore.
• How does Opus 4.7 compare to Sonnet 4.6 for my workload? Group by model in the explore.
• What was I asking when costs spiked? Filter role = user and is_user_prompt = true in the per-message table.
• How much of my usage is sub-agents vs main thread? Filter is_sidechain.

• Update pricing — edit models/claude_model_pricing.yaml. Each row is one model prefix with input/output/cache rates per million tokens.
• Add a chart — open dashboards/claude_usage_canvas.yaml, add a row with one of the available component types (bar, line, donut, heatmap, table). Save, dashboard reloads.
• Add a measure — metrics/claude_usage.yaml already has counts and cost per session/message/user-prompt; add your own SQL-aggregate measure.

The cost figures match Anthropic's actual API charge within a few percent. Known gaps:

• 1M-context Opus is billed at 2× standard rates above 200k input tokens, but the model name in the transcript (claude-opus-4-7) doesn't tell us the context size used.
• Cache write TTL — Anthropic charges 1.25× input for 5-minute writes and 2× for 1-hour writes. The transcript only stores the combined total, so we assume 5-minute (the Claude Code default).
• Subscription billing — if you're on Claude Pro/Max via OAuth, you pay the flat subscription fee, not the dollar amount shown here. This dashboard reports the API-equivalent cost (what the same usage would have cost on the pay-per-token plan).

Claude Code writes one JSONL file per session under ~/.claude/projects/<project-slug>/<session-id>.jsonl, with one event per content block. The claude_messages model reads all of them with DuckDB's read_json glob, extracts message content, tokens, and metadata, joins to a small pricing lookup table on a longest-prefix match, computes cost per row, then collapses streaming-chunk duplicates by grouping on request_id (merging text and summing block counts, taking cost once per API call). The metrics view sits on top of the model and powers both dashboards.