frugal

Route every LLM call to the cheapest model that won't compromise quality.

Frugal wraps any command with a local OpenAI-compatible proxy, classifies each request, and routes to the cheapest model that clears the quality bar. No account. No code changes. One command. 

$ curl -fsSL https://frugal.sh/install | sh
$ frugal python my_app.py
# frugal starts a proxy, sets OPENAI_BASE_URL, runs your command,
# and shuts down when it exits. Your app doesn't change.
No account No code changes Keys stay local Signed releases
View on GitHub See usage →

The insight

You're paying for capability you don't use on 60–80% of your LLM calls. A creative brainstorm doesn't need o3. A simple extraction doesn't need claude-opus. Frugal classifies each request — complexity, domain, capabilities required — and picks the cheapest model that clears the quality bar.

Classifier
Detects code, math, tool-use, JSON output, conversation depth, and domain. Combines them into a complexity score, 0.0 – 1.0.
Router
Per-quality-tier thresholds on reasoning, coding, creative, and instruction-following. Cheapest model that clears all thresholds wins.
Fallback
Routed model errors? Walk a caller-supplied chain or drop to a looser quality tier. Bounded retries so latency stays predictable.

How it works

frugal python app.py
       │
       ├─ starts proxy on a free port
       ├─ injects OPENAI_BASE_URL into your command's environment
       ├─ classifies each request (complexity, domain, capabilities)
       ├─ routes to cheapest model that clears the quality bar
       └─ shuts down proxy when your command exits

Works with any OpenAI-compatible SDK

Frugal speaks the OpenAI chat-completions API. Point your existing SDK at the proxy and nothing else changes.

Python
# unchanged from before frugal
from openai import OpenAI
client = OpenAI()
resp = client.chat.completions.create(
    model="auto",
    messages=[{"role": "user", "content": "hi"}],
)
Node
// unchanged from before frugal
import OpenAI from "openai";
const client = new OpenAI();
const resp = await client.chat.completions.create({
  model: "auto",
  messages: [{ role: "user", content: "hi" }],
});
curl
curl "$OPENAI_BASE_URL/chat/completions" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{"model":"auto","messages":[{"role":"user","content":"hi"}]}'
Go
client := openai.NewClient(os.Getenv("OPENAI_API_KEY"))
resp, _ := client.CreateChatCompletion(ctx, openai.ChatCompletionRequest{
    Model: "auto",
    Messages: []openai.ChatCompletionMessage{{
        Role: openai.ChatMessageRoleUser, Content: "hi",
    }},
})

Three quality tiers — pick per request

Send X-Frugal-Quality: cost for a quick extraction. Send high for your agent's planner step. Default balanced for everything else.

high

Top-tier models only. Use for planners, complex reasoning, novel code.

X-Frugal-Quality: high

balanced default

Best cost-quality tradeoff. The right choice 80% of the time.

X-Frugal-Quality: balanced

cost

Cheapest viable model. Classification, extraction, simple summaries.

X-Frugal-Quality: cost

Models routed today

Pricing synced from models.dev on every startup. Add more by editing ~/.frugal/config/models.yaml.

OpenAI

GPT-4o · GPT-4o-mini · GPT-4.1 · GPT-4.1-mini · GPT-4.1-nano

Anthropic

Claude Opus 4 · Claude Sonnet 4 · Claude Haiku 3.5

Google

Gemini 2.5 Pro · Gemini 2.5 Flash · Gemini 2.0 Flash

Install

One binary, ~10 MB. Detects your API keys. Adds frugal to your PATH. Release artifacts are signed with cosign; the installer verifies the checksum before moving the binary into place.

$ curl -fsSL https://frugal.sh/install | sh

From source

$ git clone https://github.com/brainsparker/frugal.git && cd frugal && make build

Run as a server

$ export FRUGAL_AUTH_TOKEN=$(openssl rand -hex 16)
$ frugal serve
$ export OPENAI_BASE_URL=http://localhost:8080/v1

When binding outside 127.0.0.1, Frugal refuses to start without an auth token (override with FRUGAL_ALLOW_UNAUTH=1). Prometheus metrics are served at /metrics. Full env-var reference in the README.

Built to self-host

Your keys stay local
Provider API keys never leave your machine. Frugal reads them from the environment and forwards requests upstream; there is no control plane.
Signed releases
Every binary is cosign-signed (keyless, GitHub OIDC). The installer verifies the checksum and, when cosign is present, the signature before executing anything.
Observable by default
Structured logs, X-Request-ID propagation, Prometheus /metrics, and /v1/routing/explain to see exactly which model was picked and why.
Source-available, BUSL 1.1
Self-hosted and internal commercial use is permitted. Each version converts to Apache 2.0 four years after release.