Emberline

Local inline code completion for VS Code. Ghost text from a code model running on your own machine — no account, no telemetry, no code leaving your laptop.

On Apple Silicon, it sets itself up

On an Apple Silicon Mac the extension bundles the llama-server engine and, the first time you type in a code file, offers to finish setup for you: it installs a small local inference server and downloads the model (~1.6 GB, once). You approve one prompt and wait for the download — no terminal, no Python, no Homebrew. Everything runs on your machine and nothing is sent anywhere.

On other platforms the extension still brings its own server, but you provide llama.cpp yourself. See Setup for both paths, and for running the server yourself if you'd rather.

Why

Every hosted completion product sends your buffer to someone else's machine and lands at 690 ms or worse at p75. If you're going to wait anyway, you may as well wait locally and keep the code. Emberline runs llama.cpp with a Qwen2.5-Coder FIM model and gets a completion back in ~950 ms warm on a base M1 — the slowest Apple Silicon chip llama.cpp lists — and ~3 ms on a cache hit.

Requirements

• macOS 15+ (Sequoia) on Apple Silicon (Metal), or Linux with a Vulkan-capable GPU (NVIDIA, AMD or Intel). Both are tested.
• VS Code 1.104+
• ~2 GB of disk for the model, downloaded on first run
• An internet connection for that first-run download

Nothing else on Apple Silicon (macOS 15+): the inference engine is bundled and the server ships inside the extension. On other platforms, and on macOS 14 or older (which cannot load the bundled engine), you also need llama.cpp — see below.

Setup

Apple Silicon — nothing to do. Install the extension, open a code file, and start typing. The first keystroke offers to finish setup; approve it and wait for the model to download (~1.6 GB, once — the status bar sits at "starting" for a minute). After that, ghost text appears after about a second; Tab accepts it. The status bar shows Emberline's state and the latency of the last completion — click it to toggle completions off and on.

If the status bar shows a warning, run Emberline: Show Logs from the Command Palette. If setup itself failed, Emberline: Show Server Log has the server's own account of why.

Linux and Windows — install llama.cpp first

Only the Apple Silicon build bundles an inference engine. Elsewhere, put llama-server on your PATH (or point EMBERLINE__LLAMA_BINARY at it) before running setup — otherwise setup stops and tells you exactly this.

Grab a build from llama.cpp releases. On Linux the Vulkan build is usually the right pick: it drives NVIDIA, AMD and Intel GPUs with no CUDA or ROCm toolchain to install.

If your machine has more than one GPU — any laptop with integrated graphics beside a discrete card — check llama-server --list-devices and pin the fast one:

mkdir -p ~/.config/environment.d
echo 'EMBERLINE__LLAMA_EXTRA_ARGS=--device Vulkan1' > ~/.config/environment.d/emberline.conf
# log out and back in

llama.cpp otherwise splits the model across every device it finds, including the slow one. Measured on an RTX 4060 laptop: 49 tok/s split, 80 tok/s pinned.

Running the server yourself

If you'd rather manage the server (another machine, a shared box, a tuned configuration), set emberline.manageServer to false and start it yourself:

# The server ships inside the extension; run it yourself with any Node 22+.
node ~/.vscode/extensions/l0kifs.emberline-*/dist/server.js
curl -s http://127.0.0.1:8011/health   # {"status":"ok",...}

Point emberline.endpoint at it if it isn't on the default http://127.0.0.1:8011. With manageServer off, the extension only ever connects to a server you started — it never installs or spawns one.

Settings

The server is configured separately, through EMBERLINE__* environment variables. See the server configuration docs.

Privacy

Your code goes to emberline.endpoint and nowhere else. By default that is 127.0.0.1, a local server running a model on your own GPU. The extension sends no telemetry.

When Emberline sets the server up for you, the only thing it downloads is the model, from Hugging Face. The server itself ships inside the extension and runs on VS Code's own Node, so there is no toolchain to install. Everything downloaded lands under ~/.emberline; rm -rf ~/.emberline removes it. None of this sends your code anywhere.

Two more details worth knowing. The extension sends the paths of your other open files (not their contents) so the server can build cross-file context; turn that off with emberline.sendOpenFiles. And completions you accept are stored locally in ~/.emberline/examples.jsonl and reused as few-shot examples — delete the file or set EMBERLINE__EXAMPLES_ENABLED=false to opt out.

Because emberline.endpoint decides where your buffer is sent, it's marked as restricted in untrusted workspaces — a malicious repo can't silently repoint it.

Commands

Known limitations

• The model is a code model. It completes code well in any language, including fenced code blocks in Markdown — but prose in Markdown or plain text is noticeably weaker and sometimes runs on repetitively. Those languages are enabled by default anyway, because a new unsaved file is plaintext until you pick a language; add them to emberline.disabledLanguages if the prose suggestions get in your way.
• Apple Silicon and Linux/Vulkan are the tested platforms. Windows should work with llama-server on PATH, but is untested.
• Only the Apple Silicon build bundles an inference engine; elsewhere you install llama.cpp yourself.
• Cross-file context is read from disk by the server, so it doesn't work in virtual or remote workspaces.

Links

• Source, architecture notes, and benchmarks
• Issues

MIT licensed.