frugon

Free, local, open-source LLM cost analyzer

TLDR

Analyze LLM usage logs and see potential savings by routing to cheaper models

$ frugon analyze ./logs.jsonl

Capture LLM calls locally via a proxy shim to generate logs

$ frugon capture --out ./logs.jsonl

Run a demo analysis on bundled sample data

$ frugon analyze --demo

List available models for routing candidates

$ frugon models

Generate a shareable HTML or Markdown report of the analysis

$ frugon analyze ./logs.jsonl --report savings.html

Install via pipx for permanent use

$ pipx install frugon

Frugon is a local-first tool that analyzes your real LLM API call logs (in OpenAI-compatible JSONL format) to show exactly where your spend is going and how much you could save by routing easier calls to cheaper models while keeping hard calls on capable ones.It runs entirely on your machine: no logs or keys are sent anywhere. It uses tokenizers and current pricing data locally. Optional `--measure` mode samples your traffic against candidate models using your own provider keys (directly, never via Frugon) to give quality estimates.Frugon can act as a transparent local proxy (`frugon capture`) that logs calls while forwarding them unchanged to the real provider.

PARAMETERS

analyze

`--candidates` comma list of models to consider for routing

`--measure` sample real traffic for quality comparison (requires `pip install 'frugon[measure]'` + provider keys)

`--judge` score candidate quality when used with `--measure`

`--report` path to write .html or .md report

`--demo` use bundled sample instead of a file

`--verbose` extra accounting detail

capture

`--out` output JSONL file (default capture.jsonl)

`--port` listen port (default 8787)

`--upstream` override the forwarding target

`--proxy` opt in to ambient HTTP(S)_PROXY for upstream calls (ignored by default so keys stay direct)

`--verbose` log one line per captured call

COMMANDS

analyze [file]

Analyze a JSONL log file (or `--demo`) and recommend routing.

capture

Run a local proxy that logs requests/responses to a JSONL file.

models [filter]

List models available for `--candidates` with pricing/quality tiers.

update

Refresh local pricing and quality data.

pricing, quality

Inspect local pricing and quality datasets.

CAVEATS

- Logs must be in the specific JSONL shape (model + request/response + optional usage).- `--measure` makes real API calls to your providers and incurs their cost.- Quality tiers are estimates; always verify with `--measure --judge` on your data before switching production traffic.- Routing recommendations are offline math; they do not automatically change your application.

RESOURCES

Source code · Homepage · Documentation