Niluxv1.0.11
v1.0.11Memory + multi-agent now in public beta

An AI engineer that lives in your terminal

Nilux reads, writes, runs, and remembers — across sessions and across agents. Built for developers who never leave the command line.

Read the docs macOS · Linux · WSL
~/projects/my-app
standard · model0 tokens
scene/refactor
nilux — zsh — 142×38
session · b94f2a
nilux agent active
memory 34 / 1M
mcp github · linear · postgres
p50 142ms
Cache hit rate97%real production avg
Cost per 1M input tokens$0.28Standard · no plans, no seats
Cost per 1M cached tokens$0.0056Standard · 50× cheaper
Context window1M+ persistent memory
/features

Built like a Unix tool, not a chatbot.

One binary, four ideas. Composable with everything you already use — pipe it, script it, automate it, forget it's there.

01 · Memory~/.nilux/memory.db

It remembers your project. So you don't have to re-explain it.

A local-first context store that survives reboots, branches, and entire sprints. Pin facts, attach decisions, scope per-repo.

projectmy-app· 12 days old
stacknext.16 · trpc · prisma
stylearrow fns · no default exports
avoidtouching billing/* without flag
memory
34 / 1M
02 · Agents→ delegate

Spawn agents. Watch them work.

Hand off a task and keep coding. They report back when they're done.

refactor
84%
tests
64%
migrate
30%
docs
03 · Tools17 built-in

Read. Edit. Search. Run.

Every tool you'd give a junior engineer, already wired up. No plugin marketplace.

▤ read
✎ edit
▸ grep
$ bash
⌥ web
❖ memory
⌥ git
⌥ +10
04 · MCPmodel-context-protocol

Plug in everything else.

First-class MCP support — connect GitHub, Linear, Postgres, your internal services. Discovered tools become available to every agent automatically.

github
linear
postgres
sentry
slack
notion
+ yours
···
/how-it-works

Three commands. Then it's just you and your terminal.

No GUI. No web dashboard. No “open in browser to continue.” Auth happens in your shell, configuration is a single JSON file, and everything streams back as text.

01 · InstallOne npm install, then it's on your PATH.No SDK, no daemon. One global npm install, ~13 MB.
02 · ConnectPair your account in 4 seconds.One device-code flow. Your tokens live on disk; we never store your code.
03 · CodeTalk to it. Or pipe to it.Use it interactively, or invoke it from scripts, hooks, and CI.
$ npm install -g @nilux-ai/code
→ fetching [email protected]
→ added to PATH · 9.1 MB · 1.2s
✓ ready · run 'nilux' to launch
$ nilux
→ not signed in — opening browser to authorize…
waiting for confirmation…
✓ signed in as @you · pay-as-you-go
$ nilux
❯ add rate-limiting to /api/upload
› read · grep · edit · test
3 files changed · 22 lines
✓ done · 6.4s
/pricing

Pay for what you ran. That’s it.

No seats, no plans, no minimums. Top up $10 and use it for as long as it lasts.

Per-token pricing.

Same model. Same context. Same tools. You see exactly what every prompt cost — to the thousandth of a cent.

$0.28/1M input tokens
$0.56/1M output tokens
Quick fix (1 session)≈ $0.03
Feature build (2–3 sessions)≈ $0.08 – $0.17
Active coding day≈ $0.10 – $0.30
Heavy coding day (600+ calls)≈ $0.50 – $1.50
Monthly (21 workdays)≈ $5 – $30
// based on 6,800+ real production calls · standard model
/voices

What developers say after a week.

Feedback from developers who made Nilux part of their daily terminal workflow.

My first AI coding agent. Nilux helps me a lot. There's no monthly subscription, you simply pay as you use it, and the price is really affordable.
AL
Alex
front-end · freelance
Nilux is a genuinely interesting product that makes my work a lot easier — the AI here isn't just another chatbot, but a real assistant inside my working environment. What I like most is the approach to safety. Nilux can analyze, suggest solutions and act as an assistant, but the final decision always stays with me: I keep full control over every file change and every command. Another strong side is memory between sessions, which makes work incredibly easier and spares me from explaining the context from scratch every time. And a separate plus — a transparent payment model with no subscriptions and no limits, where you pay only for what you actually use.
AP
Anastasiia P.
web developer
This AI is genuinely top-tier, especially for the money. I use it together with Claude at $20, and the two complement each other perfectly. I do part of the hard tasks on Claude, and when the limits there burn out I simply switch over to this one — and I notice no drop in my work at all. On top of that, it's a big help with everyday PC tasks: quickly looking something up, advice on software, rewriting a text or throwing together a simple script. All in all, a daily workhorse — and one that runs on usage rather than a subscription. Recommended.
SS
Sergey ShocKKKKKKK
fullstack
Nilux AI is the first AI assistant I'm genuinely comfortable pairing with. It doesn't rewrite the whole project its own way — it respects the existing architecture. It always studies the context and the neighboring components before making a targeted edit. No surprises after a commit. And it holds context excellently. The speed is unreal — parallel code analysis turns half a day of work into an hour. With its help I easily scaled a project from simple forms to a complex app with 3D graphics. If you need an assistant that doesn't impose its own vision and speeds development up several times over — recommended.
FD
Frontend dev
/get-started

Your terminal already knows what to do. Now it has help.

Works from any directory. Remembers your project between sessions.

Read the docs
/security

Your code stays under your control

How Nilux handles your code — and keeps you in charge of every change.

Runs on your machine

File reads, edits, and command execution all happen locally. The server decides what to do; your machine does it.

Nothing without your approval

Every edit is shown as a diff you approve or reject, and every command is shown before it runs. Four permission modes let you set how much it asks.

Dangerous commands always stop

Destructive commands hard-stop for confirmation even in auto mode, and deletes carry extra protection.

Your code context is used only to answer you

It's sent to the model to process your request, held in memory for the session, cleared when it ends, and never permanently stored on our servers.