@sxnalabs/burnlog

Private leaderboard for AI token burn. Track every token you push through Claude Code, Codex, Cursor, Gemini, Copilot, pi, and 50 other AI coding agents — tokens only, never prompts.

npx @sxnalabs/burnlog

One command. It scans for your agents and shows you your own numbers first, then signs you in through the browser, uploads, and offers to auto-sync from then on. Nothing is asked of you before you've seen what it found.

Two ways it counts

1. Log scraping — burnlog opens local agent log files and extracts their usage blocks. It uploads token counts, model, timestamp, and an opaque dedup ID; it does not upload prompts, code, filenames, or project paths.

| Agent | Where the numbers come from | | --- | --- | | Claude Code | ~/.claude/projects/**/*.jsonl | | OpenAI Codex | ~/.codex/sessions/**/*.jsonl | | Cursor | Cursor's own usage records, fetched with the local Cursor login (sent only to cursor.com) | | Gemini CLI | ~/.gemini/tmp/** | | GitHub Copilot | ~/.copilot/ (CLI + desktop) and VS Code chat sessions | | opencode, pi, Amp, Factory Droid, Cline, Roo Code, Kilo Code, Goose, Zed, Qwen Code, Kimi, Junie, Kiro, Antigravity, Grok, Hermes, Devin, Augment, jcode, Codebuff, and ~30 more | each agent's own session store — burnlog scan lists what it found | | open sink | ~/.burnlog/events/*.jsonl |

Every source's root can be overridden with BURNLOG_<SOURCE>_DIR. Session parsing is ported from tokscale (MIT).

2. burnlog wrap — for tools that keep no usage record at all (aider, Windsurf, your own scripts), count at the wire instead:

burnlog wrap -- aider
burnlog wrap -- python my_agent.py
burnlog wrap --only anthropic,openai -- npm run agent
burnlog wrap --list

wrap runs your command with the standard base-URL environment variables pointed at a loopback proxy, forwards each request upstream untouched, and reads the usage block out of the response — streaming or not.

Supported: Anthropic, OpenAI, Google, Mistral, Cohere, OpenRouter, Groq, xAI, DeepSeek, Together, Fireworks, Perplexity, Cerebras, Ollama. Anything else with an OpenAI-compatible API works via OPENAI_BASE_URL.

There is no HTTPS interception and no local CA — that's why it's base-URL redirection rather than a CONNECT proxy. Request bodies are never inspected beyond reading model when the response doesn't name it. Events are written to ~/.burnlog/events/ before any upload, so a wrapped command keeps working with no network and no API key.

The open sink

Any tool, in any language, can participate without waiting on a burnlog release. Append one JSON object per line to ~/.burnlog/events/<date>.jsonl and burnlog sync picks it up:

{"model":"gpt-5","inputTokens":1200,"outputTokens":340,"source":"my-tool","timestamp":"2026-08-11T14:00:00Z"}

Or use the command:

burnlog log 42000 --model claude-opus-4-6 --source my-batch-job
burnlog log 0 --in 30000 --out 12000 --model gpt-5

Commands

burnlog                       scan, sign in, sync, install the hook (same as `setup`)
burnlog login [api-key]       browser sign-in; pass a key for CI (~/.burnlog/config.json, mode 600)
burnlog scan                  parse logs locally, show totals (does not upload)
burnlog scan --json           exact per-source token buckets; no paths or session IDs
burnlog sync [--quiet]        upload new burn events to the leaderboard
burnlog wrap -- <cmd>         count any command's LLM calls, any provider
burnlog log <tokens>          record usage by hand
burnlog challenge             list challenges · `new` to start one · `join <code>`
burnlog status                show sync health, active sync mechanisms, and rank
burnlog budget                show joined team budget usage
burnlog budget --club <slug>  show one team by slug or id
burnlog budget --fail-on-over exit 2 when any joined team is over budget
burnlog report [--club <slug>] export team usage CSV
burnlog install               install a 30-minute schedule and Claude Code hook
burnlog install --schedule-only  install the schedule without a hook
burnlog uninstall             remove both schedule and hook; keeps sign-in
burnlog daemon                sync every 30s while this process runs
burnlog logout                delete the stored api key

CI budget gate:

BURNLOG_API_KEY=blg_team_or_user_key burnlog budget --fail-on-over --json
BURNLOG_API_KEY=blg_user_key burnlog budget --club platform --fail-on-warning

Use a team API key to check only that team. User API keys check every team the user has joined.

CSV report:

BURNLOG_API_KEY=blg_team_or_user_key burnlog report --club platform --from 2026-06-01 --to 2026-06-30 --out burnlog.csv

With a team API key, --club can be omitted because the key sees only one team. Service usage in the CSV is split by team API key label for new events.

Auto-sync and recovery

burnlog install creates a user-level 30-minute cron job (macOS/Linux) or Task Scheduler task (Windows). It also adds a Claude Code session-end hook for faster updates. --schedule-only skips the hook. burnlog uninstall removes both; it does not sign you out. Neither install nor connect is needed for manual sync.

burnlog status shows the installed mechanisms, last attempt, last success, and last error. A quiet scheduled run still exits nonzero on upload failure and saves the error for the next status check. A failed batch leaves the sync checkpoint unchanged, so the next run retries it. burnlog sync --full scans all history; already uploaded request IDs are skipped by the server.

Windows PowerShell command discovery

If npm install -g @sxnalabs/burnlog succeeds but burnlog is not recognized, use npx -y @sxnalabs/burnlog@latest status (or replace status with connect, sync, or another command). Check the global executable directory with npm prefix -g; on Windows it should contain burnlog.cmd. Get-Command burnlog tells you whether PowerShell can find it. If the shim exists but the command is missing, add that directory to your user PATH, then open a new PowerShell session. If burnlog.cmd does not exist, reinstall the package. A new session is needed after changing PATH, not merely after installing into a directory already on PATH.

Reconcile token totals with ccusage

Run burnlog scan --json and compare its exact rankedTokens per source with ccusage over the same machines and full available history. Burnlog's ranked total is fresh input + output + cache creation; cache reads are reported in a separate bucket and excluded. Codex's input_tokens includes cached_input_tokens, so Burnlog subtracts cached input before ranking. Claude Code's cache read and cache creation fields are separate from input_tokens; creation counts toward rank. The adapters also differ in which source files and agents they support. scannedFiles and events help identify coverage gaps without sharing private logs.

Claude Code nested subagent transcripts are included from 0.9.2. To backfill newly discovered events after upgrading, run burnlog sync --full once. This does not rewrite older server events with the same request ID; historical totals created by older adapter versions may still differ. If a gap remains, share only scan --json aggregates and source/version details, never transcripts or session IDs.

Environment

BURNLOG_API_URL        override the leaderboard url (default https://burnlog.net)
BURNLOG_API_KEY        api key for CI/non-interactive use
BURNLOG_CLAUDE_DIR     override ~/.claude/projects
BURNLOG_CODEX_DIR      override ~/.codex/sessions
BURNLOG_HERMES_DIR     hermes log dir (once it writes usage)
BURNLOG_OPENCLAW_DIR   openclaw log dir (once it writes usage)

Self-hosting

burnlog is open source. You can run the entire stack — Postgres + Next.js — and point the CLI at your instance:

BURNLOG_API_URL=https://burn.mycompany.com burnlog sync

Source: github.com/sharziki/burnlog.

Privacy

What we store: token counts, model name, provider, timestamp, a random dedup id. What we don't: prompts, responses, filenames, cwd, session ids, tool output, anything content-like. Full details: burnlog.net/privacy.

License

MIT © Sharvil Saxena / SXNA Labs