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.
The native AI IDE for macOS
AI coding that
respects your machine.
My MacBook Pro M4 has 16GB of RAM. Cursor and Docker were using all of it: YouTube started stuttering, and I was swapping like it was 2009. So I built Compass: a Swift-native IDE with agentic coding, on-device intelligence, and zero Electron tax. Same future. Your machine, your rules.
<100ms
On-device FIM completion (target)
200MB
Full app, AI included
0
Bytes egress until you opt in
20+
Agent tools, your provider
MIT
Open source, auditable
Sound familiar?
You didn't switch editors because the old ones were bad.
You switched because AI changed what an editor could be. This is where that story went, and where it goes from here.
- You installed Cursor because it felt like the future.
It is, and you were right to want that. Agentic coding is the biggest shift since the text editor itself.
- But your Mac pays for it.
The editor alone eats a gigabyte. Add Docker, Chrome, Slack: a 16GB machine starts swapping, fans spin up, YouTube stutters. You're paying the RAM tax for a text editor.
- And every keystroke ships your code to someone else's servers.
$20/month per seat, whether you use it or not. The usage page gets less transparent. The roadmap belongs to a board now.
- So what if the AI IDE ran native on Apple Silicon?
~200MB. SwiftUI, AppKit, Liquid Glass. Sub-100ms completion from a model that lives on your machine, not in a data center.
- What if agentic power ran through your provider, your key, your bill?
Planner → Worker → QA orchestration with 20+ tools, routed through OpenRouter, Kilo, DeepSeek, or any OpenAI-compatible endpoint. No per-seat tax.
- And what if it kept working at 30,000 feet?
No network? No problem. RAG over your whole codebase, FIM completion, local chat: the intelligence is on-device too.
- What if it were MIT open source?
Audit it. Fork it. It can't be sold out from under you. And installing it is one command.
Architecture
Two pipelines. One editor. No compromises.
Local and cloud are completely separate AI pipelines that share the same native editor. Each is tuned for what it does best, and neither pretends to be the other.
Local pipeline<100ms
Always on. Always private. No network required.
- Real FIM completion, intelligent ghost text in <100ms
- Inline AI popover, cursor-anchored Q&A, on your machine
- Semantic search, HNSW ANN, 10–50× faster, powered by the Neural Engine
- Private local chat: ask, bounce ideas, get a second opinion
- Model: 4B on-device LLM via MLX, accelerated by Apple Silicon
- No orchestration, no tool loop, by design
Cloud pipelineAgentic
On demand. Full reasoning. Your keys, your choice.
- Orchestration graph: Planner → Worker → QA → Final
- Tool loop with stall detection + recovery
- RAG context injection with evidence fusion ranking
- 20+ tools, file I/O, terminal, search, project ops, built-in browser
- Any provider: OpenRouter, Kilo, DeepSeek, Alibaba, or any OpenAI-compatible endpoint
- Opt-in only: nothing leaves your Mac until you say so
4B
Our on-device model is 4 billion parameters, running via MLX on Apple Silicon. Big enough to be genuinely useful, small enough to stay instant and private, and we never ask it to do cloud work.
Privacy by architecture
Built for the work you can't talk about
Most "AI" editors silently ship your source to someone else's servers before they'll autocomplete a single line. Compass never does. Completion, retrieval, and chat run locally; and cloud agentic work happens only when you explicitly reach for it.
Your IP stays put
- ✧ Code on a plane, in a tunnel, inside an air-gapped lab: AI still works
- ✧ Nothing leaves the endpoint without your explicit say-so
- ✧ Privacy by architecture, not by a checkbox buried in a policy
- ✧ Your provider, your key, your bill: no vendor in your loop
The assistant never forgets
- ✧ On-device RAG embeds your codebase and work history
- ✧ Ask about code you wrote last quarter; it's all still there
- ✧ Retrieval runs on the Apple Neural Engine: silent, private, instant
- ✧ Months of context survive, no dropped memory, no re-explaining
Why native wins
Native means something here
Compass is a real macOS binary, compiled in Swift and accelerated by the Metal and Neural Engine frameworks. No Chromium. No Electron event loop burning your CPU. No gigabytes of JavaScript parked in RAM. Just an editor that feels like part of the OS.
~200 MB, AI included
The entire app, with every AI capability switched on, fits in roughly 200 MB. Smaller than most editors manage with AI turned off.
No Electron tax
Tired of an IDE eating 80% of your RAM and slowing to a crawl? No JS loops, no listeners, no unaccelerated UI rendering. Just native performance.
Liquid Glass & deep OS access
Full access to low-level Apple OS features, including the latest Liquid Glass framework, something a web wrapper can never touch.
What you tolerate vs. what you get
The status quo vs. Compass
We don't need to name names. If you're using a cloud AI editor, you already know this table from the inside.
Need
The cloud-editor status quo
Compass
Autocomplete latency
500–2000ms, network-bound
<100ms, on-device
Working offline
Dead without a connection
Full offline tier: RAG, FIM, local chat
Where your code goes
Vendor's server, every keystroke
Your Mac, unless you opt in
Embeddings & RAG
Cloud-bound, code ships out
On-device, Apple Neural Engine
Agentic tasks
Strong, but locked to their models and pricing
Same orchestration with your provider, your key, your cost
Mac integration
Electron, a browser in a window
Native SwiftUI + AppKit, Liquid Glass
Memory footprint
1GB+ and climbing
~200MB, AI included
Cost & ownership
$15–20/mo per seat, roadmap owned by a board
Free. MIT. BYO keys. Nobody can sell it out from under you.
Capabilities
Packed. And shipping.
These features are in the current open-source build. No waitlist, no invite, no gate.
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.
On the roadmap
What's next, 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.
Under the hood
Engineered for the silicon it runs on
Apple Neural Engine
Embeddings and retrieval run on the ANE, so indexing and search stay silent and painless, never stealing cycles from your build.
On-device LLM
Apple MLX running a quantized 4B model. Instant, private, offline. Swap models anytime without a website update.
HNSW - FAISS
Hierarchical Navigable Small World index over FAISS vectors. 10–50× faster than brute-force with ~95–99% recall.
SQLite Index
Symbol and content search over indexed files and chunks, combined with vector retrieval for grounded answers.
Tree-sitter - Neon
Tree-sitter AST parsing with the Neon runtime for fast, accurate syntax highlighting across 20+ languages.
SwiftUI AppKit
Native NSTextView editor with SwiftUI overlays and Liquid Glass. No Chromium. No web view. Just macOS.
Fair questions
Asked and answered, in the open
The questions everyone asks before switching. The answers are the same ones we give on GitHub, HN, and Reddit.
Is the agent as good as Cursor's?
Agentic quality is the model you bring. Compass gives you the orchestration (Planner → Worker → QA with 20+ tools and stall detection), and you choose the brain: OpenRouter, Kilo, DeepSeek, Alibaba, or any OpenAI-compatible endpoint. Bring the best model in the world; we run it at native speed. Our orchestration is open and harness-tested, and the tests are in the repo.
It's not notarized. Is it safe?
MIT licensed, ~40k lines of Swift, CI green, and you can build it from source if you want to verify. Compass ships without Apple's paid notarization, like many open-source Mac apps; on first launch, right-click the app, choose Open, and confirm. We'd rather be transparent than pretend.
One maintainer. Will this die like other projects?
Compass is the editor I use every day to build itself. It's MIT, so even if I vanished tomorrow, it's yours, forever. I'm not planning to vanish: the roadmap is public, the code is open, and the community decides what comes next.
Why macOS-only?
Deliberately. Going deep on Apple Silicon is how the app stays ~200MB with sub-100ms completion and Neural Engine acceleration. There are plenty of editors that work everywhere. There's one native AI IDE for macOS.
What about native Anthropic/OpenAI support?
On the public roadmap. Until it lands, OpenRouter gives you Claude and GPT through a single key, today.
Local models are weak, aren't they?
We never ask local to do cloud's job. Local handles what must be instant and private: completion, Q&A, search, chat. Cloud handles what needs frontier reasoning: agentic work with your key. Right tool for the job, always.
Where are the benchmarks?
Live. Semantic search p95 is 0.46ms on a 2,000-vector index, 2,000 entries add in 158ms, embeddings run at 3–15ms on-device. Full table, methodology, and regression gates on the Benchmarks page, and you can run ./run.sh benchmark yourself. FIM latency with the 4B model is next, and we won't publish it until we've measured it.
I have years of muscle memory in Cursor.
Keyboard is keyboard. Command palette, familiar editing, native macOS feel. Fifteen minutes and you're home. Your Mac gets a gigabyte of RAM back.
Can my team use it?
Yes. One brew line to onboard, your team's keys, usage-based cloud instead of per-seat. Zero egress by default tends to make security teams happy.
How do I install it?
One command via Homebrew: brew install --cask jtrefon/compass/compass. Or grab the DMG from GitHub Releases. Either way, the first launch is a two-second confirmation: right-click → Open. That's the whole ritual.
Install
Get Compass
Open source, MIT, free: no account, no trial, no credit card. Requires Apple Silicon and macOS 26+.
Homebrew Live
One command, works with the current release.
brew install --cask jtrefon/compass/compass
One-line installer
Downloads the latest release and checks its signature.
curl -fsSL https://raw.githubusercontent.com/jtrefon/compass/main/install.sh | bash
The installer is open source: read install.sh
Ships without Apple's paid notarization, like many open-source Mac apps. On first launch, right-click the app, choose Open, and confirm. The full source is in the repo.
Roadmap
Ambitious, grounded in shipping architecture
v1.0, shipping now
- <100ms FIM completion, on-device
- Codebase RAG + semantic search (HNSW, ANE)
- Private local chat, fully offline
- Agentic cloud pipeline: Planner → Worker → QA, 20+ tools
- Your provider, your keys, your bill
v1.5, next
- FIM benchmarks published (TTFT, tokens/sec, p50/p95)
- MCP support, with sane defaults, not a config maze
- Agent permission model: approval gates, read-only mode, sandboxed actions
- Harness expansion: tool-loop regression tests for every agent fix
- Remote sessions (SSH/SFTP)
- Project memory, learns your conventions
- Native Anthropic & OpenAI providers
- Deeper macOS integration (Shortcuts, Spotlight)
- Signed & notarized builds, funded by sponsors
v2.0, the standard
- Plugin SDK & extensions
- Parallel specialized agents with context handoff
- Benchmark reports in every release
- Custom model fine-tuning pipeline
- Team features, usage-based, not per-seat
Built because it had to be
Your machine deserves better. It's ready when you are.
Download Compass, star the repo, open an issue, or tell us what's missing. This project is built in the open, and the community decides where it goes.