01
One task = one way
The only loop is each. No while, no for, no map-vs-reduce dilemma.
# the only loop in the languagetotal <- 0each item in carttotal <- total + item.price
an AI-native backend language
One task = one way. Batteries built in. A spec small enough for an AI to read in one go.
1234567891011121314
the problem
A dozen ways to do one thing. Packages that may not exist. Tokens spent on ceremony. For an agent, every choice is a chance to be wrong.
19.7%
of AI-suggested packages don't exist.
Fluxon ships every battery in the language — nothing to install, nothing to invent.
1
way to loop. One way to bind. One way to print.
Same meaning, same shape — there is no second idiom to choose wrong.
~2,700
tokens for the complete language spec.
An AI that has never seen Fluxon reads it once and writes working code.
the language
The only loop is each. No while, no for, no map-vs-reduce dilemma.
# the only loop in the languagetotal <- 0each item in carttotal <- total + item.price
http, db, ai, auth, ws, cron, queue — part of the language. No npm, no pip, no lockfile.
use http db ai # no install, no versionshttp.on :get "/health" \req -> rep 200 {ok:true}http.serve 8080
One line to call a model. Confidence, cost and latency come back as plain data.
r = ai.json "extract the order: ${text}" {product:str qty:int}if r._.conf > 0.85 # confidence is built indb.ins "orders" r
! propagates, ?? rescues, fail answers the client. No try/catch pyramids.
user = db.one "select * from users where id=$1" [id]!name = user.name ?? "guest"fail 422 "insufficient funds" # becomes the response
versus the world
A notes API: schema, two routes, a database. Most stacks start with scaffolding and dependencies. In Fluxon, the wiring is the language.
npm install express better-sqlite3 zod … + tsconfig + boilerplate
pip install fastapi uvicorn sqlalchemy + models + session wiring
0 installs — server, schema and persistence are the language
Rough token counts for the same runnable service. Fewer tokens = more context for the real problem — and fewer places to be wrong.
methodology
Measured what AIs actually write reliably: canonical form beats choice, batteries beat packages.
Several models independently invented a language for themselves — and converged on the same ideas.
The spec went to models that had never seen Fluxon. Every gap they hit was a real flaw.
Close the gap, re-test, repeat — from URL shorteners to e-commerce, chat and fintech.
“The goal is not to outcompete existing languages. It is to be the programming language AI knows best — and likes most.”
open source · under active development