Local-first · Open source · macOS

One becomes many.

Arslan is a local-first personal AI orchestrator. You talk to one host agent; it routes work to persona spawns you raised yourself. Their prompts improve on their own — but every change passes a held-out exam, and nothing ships until you press Promote.

↓ Download for macOS View on GitHub

macOS 11+ · Apple Silicon · signed & notarized · Apache-2.0 · pre-v1 · open source

A machine, not a chat box

Not another wrapper.

A chat box talks about the work. Arslan dispatches it — and if it can do the job itself, it does; routing is not the default. Everything runs and logs on your machine, in one thread you can read end to end.

doer-first routing · every dispatch lands in the ledger

How it works

One host. As many spawns as you raise.

Spawns are drafted from a persona seed library, named by you, kept on your Spawns Ledger. Each one equips only what you grant it — and nothing more:

🔧 tools ⚙ SKILL.md packs ◎ MCP servers

Self-evolution, with an exam

Prompts improve on their own.
Changes pass an exam first.

Your spawns get better at their jobs — without ever changing behind your back. A rewrite only replaces the incumbent prompt if it wins where it counts:

0%WIN RATE REQUIRED

0+HELD-OUT TASKS, MINIMUM

blindJUDGED, POSITIONS SWAPPED

No dimension may score worse than the incumbent.
A candidate cannot win by getting longer.
The holdout is enforced in code, not by convention.
Real and synthetic win-rates are never merged into one number.
Fail → discarded, you never see it. Pass → a readable diff, in your inbox.

Promoted. Until you press it, nothing ships. That is the whole point.

The promise guard

Caught in the act,
by design.

Turn this repo into a launch deck.

“I have handed that to Deck Master — generating now, one moment.”

0 DISPATCHES

If you have used an agent, you have been lied to exactly like this. Arslan intercepts it with a deterministic check — not another model call. One rewrite allowed; if it promises again, an honest template goes out instead.

Arslan cannot tell you it did something it did not do.

The second brain

Memory that shows
when it took effect.

What your spawns learn lands in a shared brain: [[wiki-link]] notes on a live force-directed graph, retrieved by hybrid FTS5 + embedding search.

Every entry records when it took effect — and what superseded it.
Filter the graph by start time — surviving entries, not a replay of the past.
Ask it directly: “what did my second brain learn this week?”

The real client

And this is the real thing.

Everything above is brand animation. Everything below is a screenshot — the shipped app, light theme, untouched.

New orchestrator session — real screenshot, light theme

real screenshots · ships with 6 languages · 6 theme palettes · light + dark

Safety is built in, not disclaimed

Everything that can spend or write is off by default.

From the shipped Settings page, verbatim: “They are all off by default and none of them changes anything without you.”

Generated code runs network-denied under the macOS kernel sandbox — and with no sandbox, it fails closed.
A local proxy injects git credentials; raw tokens never enter the sandbox.
Deletions go to an inbox; an overwrite keeps the original and can be undone.
Your machine · your keys · zero third-party servers in the middle.

Honest limitations

Pre-v1, and it says so.

macOS 11+ on Apple Silicon only, for now.
Bring your own model API key — Arslan runs against your account.
Auto-evolution has no spend cap yet; the product itself tells you to set a hard limit at your provider first.

One becomes many.

Not another wrapper.

One host. As many spawns as you raise.

Prompts improve on their own.Changes pass an exam first.

Caught in the act,by design.

Memory that showswhen it took effect.