Inference is all you need. Everything else can be owned by you.
Models are getting cheaper and more interchangeable by the day. What's irreplaceable is the layer around them -- your context, your memory, your tools, your identity. Others keep that layer on their servers. Osaurus keeps it on your machine.
Osaurus is the AI harness for macOS. It sits between you and any model -- local or cloud -- and provides the continuity that makes AI personal: agents that remember, execute autonomously, run real code, and stay reachable from anywhere. The models are interchangeable. The harness is what compounds.
Works fully offline with local models. Connect to any cloud provider when you want more power. Nothing leaves your Mac unless you choose.
Native Swift on Apple Silicon. No Electron. No compromises. MIT licensed.
Install
brew install --cask osaurus
Or download the latest .dmg from Releases. Two builds ship per release:
Osaurus-.dmg (~70 MB) — the standard build. Pick and download a model during onboarding. This is what Homebrew and in-app updates use.
Osaurus--full.dmg (~4 GB, linked from each release's notes and hosted on Hugging Face) — the same app with Raptor 0.6 bundled. First launch installs the model into ~/MLXModels and onboarding skips the download, so you can chat offline immediately. Updates afterwards are the small build.
After installing, launch from Spotlight (⌘ Space → "Osaurus") or the CLI:
osaurus ui # Open the chat UI
osaurus serve # Start the server
osaurus status # Check status
Requires macOS 15.5+ and Apple Silicon.
Agents
Agents are the core of Osaurus. Each one gets its own prompts, memory, and visual theme -- a research assistant, a coding partner, a file organizer, whatever you need. Tools and skills are automatically selected via RAG search based on the task at hand -- no manual configuration needed. Everything else in the harness exists to make agents smarter, faster, and more capable over time.
Agents can also opt into a private local database and a single self-scheduled next run -- see Agent DB & Self-Scheduling. Local storage is plaintext by default (protected by FileVault) with opt-in SQLCipher encryption -- see Storage.
Agent Loop
Custom agents run an agent loop. Pick a working folder and the agent gets file, search, and git tools. Toggle the sandbox and it gets shell access in an isolated sandbox. The model writes a markdown todo list, executes against it, and closes out with a verified summary -- all in the same chat window. The built-in Orchestrator is the exception: it has no working folder, sandbox, browser, or computer use. For filesystem work, create or switch to a custom agent, then pick the folder there. See the Agent Loop Guide and Orchestrator Guide.
Sandbox
Agents execute code in an isolated Linux VM powered by Apple's Containerization framework. Full dev environment -- shell, Python, Node.js, compilers, package managers -- with zero risk to your Mac.
Each agent gets its own Linux user and home directory. The VM connects back to Osaurus (inference, memory, secrets) via a vsock bridge -- sandboxed but not disconnected. Extend with simple JSON plugin recipes, no Xcode or code signing required.
The Linux VM requires macOS 26+ (Tahoe). On earlier versions Osaurus falls back to a native macOS Seatbelt sandbox: commands run on your Mac under sandbox-exec confinement, with writes limited to the sandbox workspace, pip/npm installs (no apk), and all-or-nothing network access instead of per-domain allowlists. See the Sandbox Guide for configuration, built-in tools, and plugin authoring.
Memory
Three layers -- identity, pinned facts, and per-session episodes -- plus a transcript fallback. Agents distill conversations once at session end (not on every turn), score what matters by salience, and surface at most one compact slice per request based on what you're actually asking. A background consolidator decays, merges, and evicts so memory stays sharp instead of bloating. Most turns inject ~800 tokens or less; many inject zero. See the Memory Guide.
Privacy Filter
When you send to a cloud model, an on-device classifier — OpenAI's openai/privacy-filter (Apache-2.0, 1.5B params / 50M active sparse-MoE), served via the MLX conversion mlx-community/openai-privacy-filter-bf16 (~2.8 GB) — detects names, emails, phones, URLs, addresses, dates, account numbers, and free-form secrets, alongside deterministic regex for SSN, credit cards, IBAN, AWS keys, GitHub tokens, and your own custom patterns. Each detection is shown in a review sheet with a scrubbed preview before sending; approved entities are swapped for stable [PERSON_1] / [EMAIL_2] placeholders, and streaming replies are unscrubbed back on the fly so the chat reads naturally. Fail-closed: if the post-scrub scan finds anything that leaked, the send is blocked. Verify wire-level redaction in the Insights panel — it captures the exact bytes the cloud saw. See the Privacy Filter Guide.
Identity
Every participant -- human, agent, device -- gets a secp256k1 cryptographic address. Authority flows from your master key (iCloud Keychain) down to each agent in a verifiable chain of trust. Create portable access keys (osk-v1), scope per-agent, revoke anytime. See Identity docs.
Relay
Expose agents to the internet via secure WebSocket tunnels through agent.osaurus.ai. Unique URL per agent based on its crypto address. No port forwarding, no ngrok, no configuration.
Secure Channel
When two Osaurus agents talk -- across your LAN or across the world through the relay -- the conversation is end-to-end encrypted: a forward-secret X25519 handshake authenticated by each agent's crypto identity, with every request, streamed token, and access key sealed in ChaCha20-Poly1305. The relay becomes a blind pipe that forwards ciphertext it cannot open; a man-in-the-middle cannot complete a handshake; replayed or truncated traffic is detected and refused; and there is no plaintext fallback an attacker can force. Zero configuration -- pairing is all it takes. See the Secure Channel docs.
Orchestrator
The built-in Osaurus agent is the default Orchestrator. New chat windows open on it unless you start a chat on a custom agent. Ask it anything about the app and it answers from live state; ask it to change a setting and it plans the change, shows an approval card, and applies only after you confirm. It also delegates work: spawn your custom agents and allowed local/cloud models as subagents -- in parallel, within budgets you set -- and get their results woven back into the conversation. It deliberately does no hands-on work itself -- no sandbox, working folder, browser, or computer use; those belong to custom agents -- which keeps it safe and predictable. The chat composer does not offer a working-folder chip on the Orchestrator; picking a folder there would not grant file tools. For filesystem work, create or switch to a custom agent, then pick the working folder on that agent. Rename the Orchestrator, give it a persona, and control exactly what it may delegate to in Settings → Orchestrator. See the Orchestrator Guide.
Subagents (Spawn)
Let a chat -- the Orchestrator or any custom agent -- delegate a bounded task to another model -- local or remote -- or to one of your saved agents, and get a compact result back without cluttering the conversation. Perfect for offloading research, coding, or analysis to a specialist mid-turn. When the subagent runs on a different local model, Osaurus normally does a single-residency handoff: it unloads your chat model, runs the job, then reloads and continues, so two large models never fight for memory. Same-model children reuse the resident model; experimental RAM-safe coexistence can keep both resident when enabled and admitted. Custom agents are added to the Orchestrator's spawn pool automatically (you can remove them in Settings → Orchestrator). First use still asks for approval; budgets and permission mode are configured per agent.
Browser Use
Give an agent its own browser. browser_use runs a self-contained subagent that navigates, reads pages, clicks, types, and verifies each step against a persistent per-agent WebKit session — cookies and sign-ins survive between chats, isolated from other agents and your regular browser. Reads and navigation run automatically; typing and anything consequential pause for your approval, and sign-ins happen in a secure window you type into directly (agents never see credentials). Off by default, enabled per agent, sessions managed in Settings → Browser. See the Browser Use Guide.
Image Generation
Create images on your Mac, fully offline. Install a local image model -- z-image-turbo, FLUX, Qwen-Image, or Ideogram -- and generate from a text prompt with control over size, seed, and negative prompts; hand it a source image to edit instead of starting from scratch. Your chat model can also call the built-in image tool to generate or edit a picture and render it inline in the conversation. Nothing is sent to a server.
Models
The harness is model-agnostic. Swap freely -- your agents, memory, and tools stay intact.
Local
Run Gemma 4, Qwen3.6, GPT-OSS, Llama, and more on Apple Silicon with optimized MLX inference. Osaurus maintains its own optimized model library on Hugging Face with curated quantizations for the best quality-to-size ratio on Apple Silicon. Models stored at ~/MLXModels (override with OSU_MODELS_DIR). Fully private, fully offline.
Liquid Foundation Models
Osaurus supports Liquid AI's LFM family -- on-device models built on a non-transformer architecture optimized for edge deployment. Fast decode, low memory footprint, and strong tool calling out of the box.
Apple Foundation Models
On macOS 26+, use Apple's on-device model as a first-class provider. Pass model: "foundation" in API requests. Tool calling maps through Apple's native interface automatically. Zero inference cost, fully private.
Cloud
Connect to OpenAI, Anthropic, Gemini, xAI/Grok, Venice AI, OpenRouter, Ollama, LM Studio, or Osaurus Router. Osaurus Router is the hosted path — pay-as-you-go with credits and no separate provider account; Osaurus doesn't store your prompt or response content (only the usage metadata needed for billing), and upstream providers process requests under their own policies. Context and memory persist across all providers. See Remote Providers and Osaurus Router.
MCP
Osaurus is a full MCP (Model Context Protocol) server. Give any MCP-compatible client access to your tools with the command-based stdio bridge:
osaurus mcp starts a stdio MCP server for the client and proxies tool discovery/calls to your local Osaurus HTTP server. Local-only servers can rely on loopback trust. If Server > Network exposure is enabled, add an access key to the client environment:
In the other direction, Osaurus can also act as an MCP client and aggregate tools from URL-based remote MCP providers. One-tap connect to ~25 well-known providers (Linear, Notion, GitHub, Vercel, Supabase, Sentry, Stripe, Cloudflare, ...) with auto OAuth 2.1 + Dynamic Client Registration, or paste an API key. The Remote MCP Providers UI is for HTTP/SSE MCP endpoints; it does not launch third-party command/args stdio providers. See the Remote MCP Providers Guide for details.
Tools & Plugins
osaurus tools install osaurus.xlsx # Install from registry
osaurus tools list # List installed
osaurus tools create MyPlugin --swift # Create a plugin
osaurus tools dev com.acme.my-plugin # Dev with hot reload
Apple apps are built in. Calendar, Reminders, Contacts, Notes, Mail, Messages, Maps & Location, Music, and Shortcuts ship as native tools you turn on per custom agent under Agents → Abilities → Tools — off by default, with the macOS permission prompt on enable and an approval card for every send or delete. See the Apple Apps guide.
Native plugins cover the rest: Vision, XLSX, PPTX, Images, Telegram, Resend, Fetch, and more. (Web search, browsing, macOS control, and the Apple apps are core Osaurus capabilities now — no plugin needed.) Plugins target the v3 host API surface — register HTTP routes, serve web apps, persist data in SQLite, dispatch agent tasks, and call inference through any model. Older v1/v2 plugins continue to load unchanged. See the Plugin Authoring Guide.
Document attachments keep structure where the file format exposes it: CSV/TSV tables, XLSX workbooks, PPTX decks, PDF page anchors, and rich document sections are parsed through the document adapter registry before they reach the agent.
More
Skills & Methods -- Skills import reusable AI capabilities from GitHub repos or files, compatible with Agent Skills. Full Claude plugins (skills, scheduled agents, slash commands, MCP providers, and CLAUDE.md context) can be imported from any GitHub repo and managed as a single bundle. Methods are learned workflows that agents save and reuse over time. All are automatically selected via RAG search -- no manual configuration needed. See Skills Guide and Claude Plugins.
Automation -- Schedules run recurring tasks in the background. Watchers monitor folders and trigger agents on file changes.
Voice -- On-device transcription via FluidAudio on Apple's Neural Engine. Voice input in chat, VAD mode with wake-word activation, and a global hotkey to transcribe into any app. No audio leaves your Mac. See Voice Input Guide.
Shortcuts, Spotlight & Siri -- Osaurus ships App Intents, so "Ask Osaurus" and "Run Osaurus Agent" are available system-wide the moment you install -- no setup. Ask your active agent and get the reply inline, or kick off a custom agent in the background. See App Intents Guide.
Developer Tools -- Server explorer, MCP tool inspector, inference monitoring, plugin debugging. See Developer Tools Guide. For the inference scheduler, model leases, continuous-batching engine, and feature flags that tune them, see Inference Runtime.
Telemetry
Osaurus collects anonymous, aggregated usage analytics via Aptabase, an open-source, privacy-first analytics project. We collect this only to understand broad user trends and preferences (how the app is used and where people run into friction) so we can make it better. It never includes your chats, prompts, files, model outputs, or keys. There are no accounts or device profiles, so events aren't tied to you.
For the exact, exhaustive list of every event and property we capture — and an explicit list of what we never collect — see Telemetry & KPIs.
Crash reporting
Crash and app-hang reporting via Sentry is a separate, independent switch from usage analytics. Unlike analytics it's opt-out — on by default and active from launch — because crash reports carry no personal information and are what let us fix real bugs; you can turn it off anytime in Settings → Privacy → Send Crash Reports. It's limited to crash and hang diagnostics — no performance tracing, profiling, failed-request capture, network breadcrumbs, screenshots, or personal information; we drop the user object and device hostname from every event, on top of disabling PII. It needs a DSN to be configured, so like analytics it's off by default in source builds.
Local development
Telemetry is off by default in source builds: with no key, the SDK is never initialized and every event is a silent no-op, so you can build and contribute without any of this. To enable it locally:
Create App/osaurus/Secrets.xcconfig (gitignored — never commit it) with the keys you want:
APTABASE_APP_KEY = A-XX-... — your Aptabase app key (analytics).
SENTRY_DSN — your Sentry project DSN (crash reporting). Optional; omit it to leave crash reporting off. Escape the scheme slashes: an .xcconfig treats // as a comment, so a raw https://… DSN gets silently truncated to https:. Add a slash variable and reference it:
In Xcode, add Secrets.xcconfig to the project (no target membership), then under Project → Info → Configurations → Debug → osaurus set "Based on Configuration File" to Secrets.
Clean build (⇧⌘K) and relaunch.
The keys flow Secrets.xcconfig → $(APTABASE_APP_KEY) / $(SENTRY_DSN) build settings → AptabaseAppKey / SentryDSN in Info.plist. Debug builds report to Aptabase's Debug bucket (enable the Debug view on the dashboard to see them) and to Sentry's debug environment, so local testing never pollutes production data.
Compatible APIs
Drop-in endpoints for existing tools:
API
Endpoint
OpenAI
http://127.0.0.1:1337/v1/chat/completions
Anthropic
http://127.0.0.1:1337/anthropic/v1/messages
Ollama
http://127.0.0.1:1337/api/chat
All prefixes supported (/v1, /api, /v1/api). Full function calling with streaming tool call deltas. /chat/completions keeps strict OpenAI semantics -- it returns tool_calls and the client executes them, so Osaurus drops in cleanly behind harnesses that already manage their own tool loop. For server-side autonomous loops use POST /agents/{id}/run; to expose Osaurus tools to remote MCP harnesses use /mcp/tools + /mcp/call. See OpenAI API Guide for tool calling, streaming, and SDK examples. Building a macOS app that connects to Osaurus? See the Shared Configuration Guide.
CLI
osaurus serve --port 1337 # Start on localhost
osaurus serve --port 1337 --expose # Expose on LAN
osaurus ui # Open the chat UI
osaurus status # Check status
osaurus stop # Stop the server
[!NOTE]
🌐 Help translate Osaurus. We're looking for contributors to localize the app into Spanish and Japanese -- these locales are already wired up in Xcode, so you can start translating right away. See docs/TRANSLATORS.md for how to contribute and the contributor leaderboard.
