LedgerLM - AI Usage & Cost

AI usage and cost tracking for Claude Code, GitHub Copilot, Codex CLI, and OpenCode - fully local.

A VS Code extension that tracks token usage for Claude Code, GitHub Copilot, Codex CLI, and OpenCode in one place. LedgerLM is token-first, with cost shown only when the data is reliable. It includes a tabbed dashboard (Overview / Sessions / Models / Agents & Tools), time-window filtering, token/cache deep dives, cache-break impact, agent & tool usage analytics, budgets, live updates, and CSV/JSON export.

_{The LedgerLM panel: per-source cards, token & cache deep dive, latest session, budgets, daily usage chart, and a by-model breakdown. It also follows your VS Code light theme:}

100% local - no cloud, no telemetry

Your data never leaves your machine. LedgerLM reads the log files your AI tools already write to disk, computes everything locally, and stores results in a local SQLite file inside VS Code's own storage. There are no servers, no accounts, no outbound network calls, and no telemetry. Source logs are opened read-only and are never modified.

Token-first by design. AI tools do not bill the same way - Copilot meters premium-request credits, Claude Code is commonly a flat subscription, and Codex rides on a ChatGPT plan. LedgerLM prioritizes tokens and cache behavior (always exact) and shows USD only when confidence is clear. See Cost methodology.

Features

Unified dashboard

• Per-source cards - tokens, cost (with confidence), credits, and cache-hit for Claude Code, Copilot, Codex, and OpenCode side by side.
• Time window - All time / 1h / 3h / 6h / 12h / 24h / 48h / 7 days / custom date range; filters every panel.
• Source selector - a dropdown to focus on one tool (Claude Code, Copilot, Codex, OpenCode) or all at once.
• Daily chart - stacked token bars per source + a combined cost line.

Token & cache deep dive

• Token breakdown - fresh input · cached read · cache write · output, as a proportion bar with counts and %.
• Cache efficiency, tokens served from cache, cache writes, and cache breaks (when a warm prompt cache was lost). See Tokens & cache.
• By-model table - per-model token split + color-coded cache-hit; sub badge for sub-agent calls; ⚠ no price for unpriced models.

Sessions

• Latest Session panel - most recent session's stats + per-model mini-table, auto-refreshing.
• All sessions - sortable columns; click a row to expand an inline per-model breakdown. sub / ⚠ badges.

Agents, tools, skills & hooks

• Subagents - which agents ran (and how often, across how many sessions).
• Tools - per-tool call counts, error counts, and average execution time.
• Skills & hooks - skill invocations and hook executions with timing, so you can spot slow hooks.

Cost - only when trustworthy

• Copilot -> premium-request AI credits (copilotUsageNanoAiu, $0.01/credit). ≥ marks a floor when some calls predate the credit field.
• Claude Code / Codex -> ≈ API-equivalent estimate from bundled price tables (you likely pay a flat subscription, so it is labeled est; hide it with ledgerLM.showEstimatedCost).
• OpenCode -> the per-message cost OpenCode itself records (provider list price) - shown as an estimate since it is not a bill.
• — → no reliable cost; tokens are shown instead. Unpriced models are excluded from totals.

Budgets, live, export

• Budget alerts - daily/weekly USD thresholds; one VS Code warning per period; status bar shows today's tokens / ≈USD.
• Live tracking - watches all log roots (fs.watch) and refreshes within ~2s as you work.
• Export - CSV/JSON of sessions (window- and source-filtered) from the dashboard or the command palette.

Where the data comes from / where it's stored

Computed data lives in a local sql.js SQLite file under the extension's VS Code globalStorage (ai-cost.db). Source logs are never modified. See Architecture.

Privacy

LedgerLM is local-first and offline by design:

• No network. Nothing is uploaded, no analytics, no telemetry, no accounts or sign-in. The extension makes zero outbound requests.
• Read-only. Your AI tool logs are opened read-only; LedgerLM never edits or deletes them.
• Local storage only. All computed data stays in a SQLite file inside VS Code's globalStorage on your disk. Uninstalling clears it.
• No bundled secrets. Pricing tables ship with the extension; no API keys are used or required.
• Open source. Audit every line - see the repository.

Install / develop

npm install
npm test                 # vitest unit tests
node scripts/smoke.js    # end-to-end sync against your real logs, prints totals
npm run dev:vscode       # package .vsix and install into VS Code
# or press F5 in VS Code for an Extension Development Host

Open the dashboard from the activity-bar icon, the status bar item, or LedgerLM: Open Dashboard.

Settings

Docs

• Architecture — modules, data flow, schema.
• Cost methodology — how cost is computed and the confidence levels.
• Tokens & cache — token types, cache breaks, and the relationship to VS Code's Agent Debug / Cache Explorer.
• Changelog.

License

MIT.