The problem
Claude Code spend arrives as one number. It doesn’t say which session, initiative, skill or agent produced it, or which work you keep repeating.
What it does
work-ledger reads one person’s Claude Code transcripts and reports usage and estimated cost at every level: session, initiative, section, turn and individual call. It attributes work by kind, tool and skill, groups sessions into initiatives, rolls up an initiative’s cost across every session it touched, and flags repeated work.
Parsing, cost calculation, rollups and the read-only web UI run locally. Grouping sessions into initiatives uses Anthropic’s hosted API by default, so short prompt snippets leave the machine; an Ollama option keeps it local. Costs come from a pricing table and are estimates.
Status
For one person inspecting their own usage, not a team observability product. The recommendations layer is early, and automation is explicitly not presented as mature. The repository’s screenshots are generated demo data, not real findings.