#2929th of 36 in Assistants

osaurus

Native macOS agent harness with memory, sandbox and tools over local or cloud models

Stars
8.1k
License
MIT
Last commit
Oct 2026
Last release
Oct 2026
Language
Swift
Stars, last 29 weeks+3,701
1 Apr 2026: 4,364 stars8 Apr 2026: 4,644 stars15 Apr 2026: 4,804 stars22 Apr 2026: 4,952 stars29 Apr 2026: 5,017 stars6 May 2026: 5,051 stars13 May 2026: 5,094 stars20 May 2026: 5,338 stars27 May 2026: 5,389 stars3 Jun 2026: 5,433 stars10 Jun 2026: 5,548 stars17 Jun 2026: 5,814 stars24 Jun 2026: 5,923 stars1 Jul 2026: 6,430 stars8 Jul 2026: 6,766 stars15 Jul 2026: 7,050 stars22 Jul 2026: 7,179 stars29 Jul 2026: 7,332 stars5 Aug 2026: 7,451 stars12 Aug 2026: 7,523 stars19 Aug 2026: 7,582 stars26 Aug 2026: 7,630 stars2 Sept 2026: 7,707 stars9 Sept 2026: 7,778 stars16 Sept 2026: 7,902 stars23 Sept 2026: 7,960 stars30 Sept 2026: 7,992 stars7 Oct 2026: 8,015 stars14 Oct 2026: 8,044 stars21 Oct 2026: 8,065 stars

4,364 on 1 Apr 20268,065 on 21 Oct 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

Osaurus is a Swift app for Apple Silicon Macs that runs local MLX models or cloud providers and wraps them with per-agent prompts, memory, tools and a code sandbox. It serves OpenAI, Anthropic and Ollama compatible APIs, acts as an MCP server and client, and includes an on-device privacy filter for cloud requests. Agents can browse, run on schedules and be reached through an encrypted relay.

Who it is for: Mac users who want self-hosted agents with memory on local or cloud models

Strengths

  • Native Swift app with MLX inference; works fully offline with local models
  • Agent code runs in an isolated Linux VM (macOS 26+) or a Seatbelt fallback
  • On-device privacy filter scrubs PII before cloud sends and blocks the send if anything leaks
  • Serves OpenAI, Anthropic and Ollama compatible APIs and acts as an MCP server and client

Weaknesses

  • macOS 15.5+ on Apple Silicon only; no Linux or Windows build
  • Linux VM sandbox needs macOS 26; older systems get weaker Seatbelt confinement
  • Crash reporting via Sentry is on by default and must be turned off in settings
  • Built-in Orchestrator has no folder, sandbox or browser; file work needs a custom agent

What it needs

  • GPU optional
  • Needs macOS 15.5+, Apple Silicon
  • Models: MLX, Apple Foundation Models, Liquid LFM, OpenAI, Anthropic

Also in Assistants

See all 36
Also in Assistants
RankProjectScore
1ai-job-searchClaude Code workflow that searches job boards and drafts tailored CVs and cover letters63.1k starsMIT92 out of 100
2OpenClawPersonal AI assistant reachable through Discord, Slack, Telegram, WhatsApp and more392.7k starsMIT85 out of 100
3nanobotSelf-hosted personal AI agent reachable from browser, terminal and chat apps49.3k starsMIT85 out of 100
#4openhumanRust agent harness with desktop app, built-in memory and embeddable library42.2k starsGPL-3.085 out of 100
#5ZeroClawSingle Rust binary agent runtime with 30+ channels and hardware access33k starsApache-2.080 out of 100