中文
← Back to Projects

Open source · pre-v1AI systems · macOS + iPhone

Arslan

Tell it once. It gets on with it.

A personal agent for the Mac: it takes the task, writes down what “done” means, works in the background and asks before it touches anything. Approve from the Mac or from your iPhone.

67
releases
2,329
commits
8,466
backend tests
Open
source, Apache-2.0

As of 2026-10-14, from the repository and releases

Arslan — Background work, with what “done” means.
01Background work, with what “done” means.Arslan's interface rebuilt from its source for the launch film; the scenario is staged.

How it works

  1. 01

    Say what you want

    Type or talk in one conversation. Arslan writes down what “done” means before it starts.

  2. 02

    It works in the background

    The job runs while you keep chatting, with its own budget, checkpoints and a Stop button.

  3. 03

    Every action passes the gate

    Commands run in a sandbox and are checked as run, ask or forbid; anything that touches your things asks first, on the Mac or the iPhone.

  4. 04

    Checked, then delivered

    Each “done when” line is checked; the job ends as Done, Partly done, Stuck or Stopped, with its files and steps.

The hard parts, and what I did

Ask once. Keep talking.

Work should leave the conversation, not hold it up: the task goes to the background, the chat stays free, and the result comes back checked.

  1. A job that stops halfway is worse than no job

    The problem
    Background jobs first shared the budget of a single chat answer and were cut off the moment it ran out. Research jobs ended with nothing to show.
    What I did
    Gave every job its own budget tier. At the tier Arslan stops looking things up and writes the result from what it already has; only a job that cannot stop itself is halted, at 1.5×.
    Result
    Shipped in 0.1.43; the card says which limit ran out and how to continue.
  2. “Done” has to be checked, not claimed

    The problem
    An agent that reports success on its own word is easy to build and hard to trust.
    What I did
    Before starting, each job writes down what “done” means. When it finishes, the criteria are checked and the result is labelled Done, Partly done, Stuck or Stopped.
    Result
    Visible on every job card; the criteria turn “met” one by one.
  3. Safety tests that never ran

    The problem
    Arslan ships on macOS, but CI ran on Linux. 25 macOS-only tests — three of them sandbox-escape checks — had never run once.
    What I did
    Added macOS as a required CI check and verified each of the four required checks actually blocks a merge when it fails.
    Result
    6,443 backend tests in CI for 0.1.55, plus 49 macOS-only cases that must run with zero skips; generated Python runs in a network-denied kernel sandbox.

Open source · pre-v1Local-first macOS app with an iPhone companion, open source (Apache-2.0) and pre-v1. Models and tools depend on your own configuration and keys.

Stack

React · TypeScript · ViteTauri · Rust desktop shellPython · FastAPISQLite · SQLAlchemy · FTS5macOS Seatbelt sandbox · Arslan Hands (Rust)iPhone: SwiftUI · CloudKit · CryptoKitLLM tool calling · MCP

Implementation details

Local-first, pre-v1. External models and tools depend on configuration.

Your context
Start with a request or an imported document.
Workspace
See replies, tool activity and decisions in one place.
One agent loop
Calls the model, runs tools behind the gate and checks the result.
Document intake
Extract text; use local OCR when a page is an image.
Local knowledge
Save records and retrieve text by full-text search.
Models & tools
Configured services provide generation and tool capabilities.