other / agent
Osaurus
A native macOS app for Apple Silicon that runs AI agents with local or cloud models, agent memory, a code sandbox and an MCP server.
At a glance
- Purpose
- A native macOS app for Apple Silicon that runs AI agents with local or cloud models, agent memory, a code sandbox and an MCP server.
- Runs in
- Local agent with shell, MCP server, Web service
- Operating systems
- macOS
- Needs
- Runtime: Not stated; Needs internet: No; Needs accounts or API keys beyond your agent: Not stated
- Files and data
- Reads files: Yes; Writes files: Yes; Sends data to other services: Yes
- Tool type
- Agent
- Delivery
- From GitHub
- Version
- Latest on the main branch
- Access
- Code on GitHub
- Compatibility
- According to the README, Osaurus runs on macOS 15.5 or later with Apple Silicon. It lists local models such as Gemma 4, Qwen3.6, GPT-OSS and Llama through MLX, Liquid AI's LFM family, and Apple Foundation Models on macOS 26 or later. The cloud providers it names are OpenAI, Anthropic, Gemini, xAI/Grok, Venice AI, OpenRouter, Ollama, LM Studio and Osaurus Router. The README describes local OpenAI-, Anthropic- and Ollama-compatible endpoints and says Osaurus works as an MCP server and client. The README says its skills are compatible with Agent Skills and that Claude plugins can be imported from GitHub repositories.
- External costs
- According to the README, Osaurus is MIT licensed and works fully offline with local models. The cloud providers it names are OpenAI, Anthropic, Gemini, xAI/Grok, Venice AI and OpenRouter. The README says Osaurus Router is a hosted, pay-as-you-go option that uses credits and needs no separate provider account. The README gives no prices. Providers you connect may charge for usage.
- Permissions
- The README shows Homebrew installation and desktop downloads, and says models are stored in ~/MLXModels unless overridden with OSU_MODELS_DIR. The README describes custom-agent file and Git tools, sandbox shell access and package managers, recurring schedules and folder watchers. The README says Apple-app tools are off by default and request macOS permissions when enabled, with approval cards for sends and deletes. The README describes persistent per-agent WebKit browser sessions whose cookies and sign-ins survive between chats. The README says reads and navigation run automatically, while typing and consequential actions pause for approval. The README lists remote MCP connections using OAuth or API keys, plugins including Telegram and Resend, and internet exposure through an agent relay. The README describes an identity system using a master key in iCloud Keychain, settings changes after approval and source-development installation of a pre-push formatting hook.
- Usage terms
- MIT License, as stated in the repository. See its LICENSE file for the exact terms. The README identifies OpenAI's openai/privacy-filter model as Apache-2.0 licensed.
- Interface
- desktop
Code on GitHub
osaurus-ai/osaurus
MITOwn your AI. The native macOS harness for AI agents -- any model, persistent memory, autonomous execution, cryptographic identity. Built in Swift. Fully offline. Open source.
Topics: anthropic apple-foundation-models apple-intelligence apple-neural-engine llm mcp mcp-server mlx
Public GitHub data fetched · 0 days old
View on GitHub ↗What it does
The README describes Osaurus as a Swift AI harness for macOS on Apple Silicon, using local or cloud models. The README says agents have prompts, memory and themes and that RAG search selects tools and skills.The README describes custom-agent file, search and Git tools, a code sandbox, layered memory, a cloud-send privacy filter, agent identities, an internet relay, browser use, local image generation and an Orchestrator that delegates to subagents.
The README says Osaurus acts as an MCP server and client and exposes OpenAI-, Anthropic- and Ollama-compatible local endpoints. The README describes tool-plugin installation from a registry and skill imports from GitHub repositories or files.
Who it is for
Mac users with Apple Silicon who want AI agents that run on their own machine with local or cloud models.Example tasks
Create a custom agent and pick a working folder for it.Turn on the sandbox for an agent so it can run shell commands in isolation.
Connect an MCP client to Osaurus through the osaurus mcp stdio bridge.
Install a tool plugin from the plugin registry with the osaurus tools command.
Import a skill from a GitHub repository or a file.
Getting started
According to the README, install with brew install --cask osaurus, or download the .dmg from the project's Releases page. The README says to launch Osaurus from Spotlight or use osaurus ui to open the chat UI, osaurus serve to start the server, and osaurus status to check status. The README says the standard build downloads a model during onboarding, while a larger build bundles one. The README's source-build instructions say to open osaurus.xcworkspace and build and run the osaurus target. The README has the exact commands.Limitations
The README requires macOS 15.5+ and Apple Silicon; the Linux VM sandbox and Apple Foundation Models require macOS 26+. The README says earlier versions use Seatbelt confinement with workspace-limited writes, pip/npm rather than apk installs and all-or-nothing network access. The README says the standard download requires a model download during onboarding and a larger build bundles a model. The README lists Xcode 16+ and macOS 15.5+ for source builds. The README says the built-in Orchestrator has no working folder, sandbox, browser or computer use. The README says browser use is off by default. The README describes unloading and reloading models for different-model local subagents and labels concurrent coexistence experimental. The README says the Remote MCP Providers UI supports HTTP/SSE endpoints and does not launch third-party command/args stdio providers. The repository can change at any time; check its README before you install.Data practices
The README describes local-model operation and an on-device privacy filter for cloud sends, with detection previews and a post-scrub check. The README says Osaurus Router retains billing usage metadata rather than prompt or response content, and upstream providers process requests under their own policies. The README says local storage is plaintext by default, protected by FileVault, with opt-in SQLCipher encryption.The README says Aptabase receives anonymous, aggregated usage analytics to understand broad trends, preferences and friction in app use. The README points to docs/TELEMETRY.md for the event and property list; it does not give an analytics opt-out control in the README. The README says Sentry crash and app-hang reporting is a separate switch, on by default from launch, and can be disabled in Settings → Privacy → Send Crash Reports. The README describes crash and hang diagnostics with performance tracing, profiling, failed-request capture, network breadcrumbs and screenshots disabled. The README says telemetry is off by default in source builds without configured keys.