burnrate

Stop overpaying for AI. burnrate audits your codebase for wasteful AI API usage, suggests targeted fixes, and can execute a full stack migration off expensive subscriptions.

The core insight: most devs pay $100–400/month for Cursor, Claude Max, and Copilot. The same output costs $20–40 via direct API + open-source tooling. Burnrate finds the gap and closes it.

Install

pipx install burnrate

Or with pip:

pip install --user burnrate

Quick start

burnrate auth                  # one-time setup: keys + system profiling
burnrate audit .               # scan your codebase (read-only)
burnrate fix .                 # patch the easy wins, with diff approvals
burnrate overhaul              # replace expensive subscriptions with API + open-source
burnrate status                # current-month spend vs subscription baseline
burnrate doctor                # verify the setup is healthy

Commands

burnrate auth

First-time setup. Prompts for API keys (Anthropic, OpenAI, Google, Groq, Mistral), detects your system (OS, RAM, shell, editor), and asks which paid AI tools you're subscribed to. Writes everything to ~/.burnrate/config.json. Keys never leave your machine.

burnrate audit [path]

Read-only scan. Finds:

• Expensive model choices (e.g. claude-opus-4 for a simple classifier)
• Missing max_tokens (unbounded response spend)
• Long system prompts (>1k chars) without cache_control
• Cross-language: Python (AST-based) and TypeScript/JavaScript (regex-based)

burnrate fix [path]

Applies code-level fixes. Shows a unified diff per finding and prompts y / n / skip / all (per-fix-kind batching). Add --dry-run to preview without writing. Fix types: model swap, add max_tokens, wrap system prompts with cache_control.

burnrate overhaul

Personalized stack-replacement plan. Reads your detected system + subscriptions; recommends replacements (Zed for Cursor, Anthropic API + Aider for Claude Max, Ollama + Qwen for Copilot). On approval, runs the installs (brew, pipx, ollama pull) with a single confirmation gate. Add --dry-run to see the plan without executing.

burnrate status

This month's API spend per provider, with subscription-baseline comparison. Queries the Anthropic and OpenAI Admin APIs (/v1/organizations/usage_report/messages and /v1/organization/costs). Requires admin-tier API keys — regular keys show "requires admin API key" gracefully. Subscription savings are shown regardless.

burnrate doctor

Verifies config, Python version, API key validity (via cheap GET /v1/models probes), and optional tools on PATH (Ollama, Aider, Zed). Exit 0 if all checks pass, 1 if any fail.

burnrate mcp

Starts an MCP (Model Context Protocol) server on stdio. Exposes four tools to MCP-compatible editors (Claude Code, Cursor, Zed):

• burnrate_check_file(path) — scan a single file
• burnrate_suggest_fix(code, finding_type) — return patched code
• burnrate_get_spend() — current-month spend
• burnrate_recommend_model(task_description) — most cost-effective model for the task

Register in your editor's MCP config:

{
  "mcpServers": {
    "burnrate": {
      "command": "burnrate",
      "args": ["mcp"]
    }
  }
}

Safety

• burnrate fix shows every diff before writing. There's no --yes shortcut for v0 — confirm each kind or pipe all to batch.
• burnrate overhaul shows every install command before running. Single confirmation gate; no per-step prompts during execution. --dry-run is always safe.
• burnrate audit is fully read-only.
• API keys live at ~/.burnrate/config.json. The CLI never transmits them to any third party.

If you want extra safety on fix, run it on a committed working tree so git diff is your undo.

Web companion

If you'd rather use a hosted UI, get personalized setup scripts, or track spend over time, burnrate.money has a paid tier ($12/mo or $49 lifetime) that complements the CLI.

Development

git clone https://github.com/saintbate/burnrate
cd burnrate/cli
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest -q

License

MIT.