LiveReal FIM completion
Ghost-text completion in <100ms from a 4B model running on your Mac via MLX. Not a remote round-trip. Not a fake heuristic.
Features
Everything you need. Nothing you tolerate.
The AI editor you actually wanted: agentic power on your terms, private intelligence on your machine, and performance that feels native, because it is.
Real AI, on your machine
Completion that's actually intelligent.
Ghost-text completion in <100ms from a 4B on-device model: genuine intelligence, not a fake heuristic. Retrieval runs on the Apple Neural Engine, so indexing and search stay silent and painless while you keep typing.
Offline? Still works. Plane, tunnel, coffee shop: your AI doesn't quit because it never needed the network.
Capabilities
In the current build
LivePrivate local chat
Ask questions, bounce ideas, get a second pair of eyes, entirely on-device. No API key, no network, no one listening.
LiveANE-accelerated RAG
Embeddings and retrieval run on the Apple Neural Engine, so indexing and search stay silent and painless, never hogging your CPU.
LiveCodebase intelligence
SQLite-backed index plus HNSW vector retrieval over files, symbols, and chunks: grounded answers drawn from your own code.
LiveAgentic coding, your provider
Planner → Worker → QA orchestration with 20+ tools. Route through OpenRouter, Kilo, DeepSeek, Alibaba, or any OpenAI-compatible endpoint. Your key, your bill.
LiveBuilt-in terminal & browser
Read docs and run commands without leaving the editor. One app, zero context-switching.
LiveInline AI popover
Cursor-anchored Q&A for instant explain, refactor, or ask, without losing your place in the code.
LiveSemantic search
HNSW ANN vector index with on-device embeddings, 10–50× faster than brute-force with ~95–99% recall.
Next up
On the roadmap, in the open
NextRemote sessions
SSH and SFTP integration so the agent works in your infrastructure.
NextProject memory
Adaptive rules and inspectable memories that teach the assistant your conventions per repository.
NextNative Anthropic & OpenAI providers
Direct provider integrations: on the roadmap, not vapor. Until then, OpenRouter covers them through one key.
NextPlugin SDK
Extensions for tools, languages, and workflows. The ecosystem question, answered in the open.
NextSigned & notarized builds
Funded by sponsors: the $99/year Apple fee, paid by the community, so your Mac trusts us out of the box.
Native to the core
Built for the silicon it runs on.
~200 MB. AI included.
The whole app, with every AI capability on, fits in roughly 200 MB. Smaller than most editors with AI turned off.
No Electron tax
A real macOS binary in Swift. No Chromium, no event loop burning your CPU, no gigabytes of JavaScript in RAM.
Liquid Glass & deep OS access
Full low-level Apple OS features, including the latest Liquid Glass framework, something a web wrapper can never touch.
Built-in terminal & browser
Run commands and read docs without leaving the editor. One app, zero context-switching.
Private local chat
MLX-backed models run chat, quick actions, and inline completion on your Mac. Offline. Private. No API key required.
Bring your own provider
Route agentic work through OpenRouter, Kilo, DeepSeek, Alibaba, or any OpenAI-compatible endpoint. Your key, your bill.
Honest about what it isn't
What Compass is not
Compass is not a Linux editor, not a Windows editor, and not an enterprise procurement product, not yet anyway. It's not a cloud-only SaaS, and it's not a browser pretending to be an app. We'd rather be the best native AI IDE for macOS than a mediocre one everywhere. If that's not your setup, no hard feelings. Come back when Apple Silicon is.
Try it. It's free, really.
MIT open source. No account, no trial, no credit card. Requires Apple Silicon and macOS 26+.