Get started in five minutes

FLUX sits between your app and your AI provider, removes the tokens you were about to pay for twice, and writes every request into a tamper-evident ledger you can audit yourself. No signup, no account, no key handover: your API key passes through to the provider and is never stored. Works with Claude (Anthropic), OpenAI, and any OpenAI-compatible endpoint.

Step 1 — run the proxy

One command, nothing to clone:

npx -y -p @flux-ai/sdk flux-proxy

You'll see:

FLUX proxy listening on http://127.0.0.1:8787
Ledger spools: ./flux-ledger

To switch on the Claude prompt-cache lever too (our single biggest measured saver, 37.6% on multi-turn traffic), start it as FLUX_PROXY_CACHE_TTL=5m npx -y -p @flux-ai/sdk flux-proxy.

Step 2 — point your app at it

Change one line. Keep your own API key — FLUX passes it through verbatim.

// Anthropic SDK
const anthropic = new Anthropic({ baseURL: "http://127.0.0.1:8787" });

// OpenAI SDK — also any OpenAI-compatible endpoint
const openai = new OpenAI({ baseURL: "http://127.0.0.1:8787/v1" });

Any language works — it's just a base URL. Python, Go, curl, LangChain, whatever you run today.

Step 3 — run your traffic as normal

Responses come back byte-identical, streaming included. Reduction is lossless-only by default: compacted tool schemas decode identically, removed duplicates are exact, and if anything fails your request goes through untouched. Each request appends one hash-chained event to ./flux-ledger/<your-app>-<provider>.jsonl — usage numbers, lever names and hashes. Never your prompts.

Step 4 — check our maths yourself

This is the part no other cost tool offers: the invoice is a pure function of a ledger file you hold. Verify the chain and recompute it with the same published package:

node -e '
const { readFileSync } = require("node:fs");
const { verifyLedgerChain, recomputeInvoice } = require("@flux-ai/sdk");
const records = readFileSync(process.argv[1], "utf8").trim().split(/\n/).map(JSON.parse);
const chain = verifyLedgerChain(records);
if (!chain.ok) throw new Error("chain broken at seq " + chain.seq);
console.log("chain verified:", chain.length, "records");
console.log(JSON.stringify(recomputeInvoice(records.map(r => r.event)), null, 2));
' flux-ledger/proxy-anthropic.jsonl

If FLUX saved you nothing, the verdict line literally reads "verdict": "do_not_charge" — that is the honest outcome on traffic that is already lean, and we publish it as such.

Already measuring with LiteLLM, Helicone or Langfuse?

You don't need to route any traffic to get a quote. Export your existing logs and run them through the quote tool — the verdict can be "do not pay us":

npx -y -p @flux-ai/sdk flux-quote your-export.jsonl --format litellm

What it costs

1% of verified savings, monthly — nothing else. No subscription, no seats, no minimum. Savings are measured per completed task from the provider's own reported usage; a month that saved nothing bills nothing. The measurement rules are published in full in the methodology, and every number we quote traces to a command and a commit on the evidence page.

What FLUX never does

• Never sends your prompts anywhere — analysis runs entirely on your machine.
• Never stores or forwards your API key beyond the provider you chose.
• Never claims savings you already had: your own cache hits are itemised and excluded.
• Never hides a loss: a lever that cost you money counts against us on the invoice.

Want a guided pilot?

Open a pilot request — bring an export from your existing gateway and leave with a per-segment quote from your own traffic, including the honest band for your workload shape.