Dakodeon

Run local models from your macOS menu bar — behind an OpenAI-compatible API.

Website · Install · How it works · Develop

---

Dakodeon is a tiny menu bar app. Start a local llama-server router and point any agent — OpenCode, Zed, or your own scripts — at http://127.0.0.1:8080/v1. Your agent chooses the model by id; Dakodeon shows what's loaded and manages the downloads.

It bundles no runtime and no weights. It drives the llama.cpp and hf tools already on your machine, so the app itself stays tiny.

Install

brew install --cask emin93/tap/dakodeon

Note

Requirements: macOS 14+, with llama-server and hf on your PATH.

brew install llama.cpp
pip install -U "huggingface_hub[cli]"   # provides `hf`

Features

🧭  Menu bar control	Start/stop the server and see the active model from a slim panel.
📦  Model manager	A Settings window shows each model's download status — download, cancel, delete, or reveal weights in Finder.
🔄  Selection in your agent	Clients like OpenCode select a model by id; llama-server routes to that profile and keeps one loaded at a time. The app has no model picker.
🧹  Clean shutdown	Quitting the app stops llama-server.
🚀  Native defaults	llama-server loads each model's trained context and embedded chat template.

How it works

Dakodeon launches llama-server in router mode and exposes the standard OpenAI-compatible endpoints. GET /v1/models returns the available profile ids, such as gemma4-12b-it-qat and gemma4-31b-it-qat. Chat requests route by the JSON model field, so switching models in a client like OpenCode also moves the active model Dakodeon shows in the menu.

POST http://127.0.0.1:8080/v1/chat/completions
GET  http://127.0.0.1:8080/v1/models

Model files download to the shared Hugging Face cache via hf; the app resolves the local GGUF paths and points the server at them — nothing is copied or duplicated. Sizes and LFS hashes come from hf models list <repo> -R --json. Models are shown by their Hugging Face repository — the same id clients send.

Models

Profiles are curated in code at Sources/Dakodeon/Catalog.swift. Each ModelProfile declares its weights, an optional draft / MTP model, and any extra llama-server flags. The app exposes no per-user configuration — to add a model, append an entry:

ModelProfile(
  id: "gemma4-12b-it-qat",
  weights: ModelAsset(repo: "unsloth/gemma-4-12B-it-qat-GGUF", file: "gemma-4-12B-it-qat-UD-Q4_K_XL.gguf"),
  draft:   ModelAsset(repo: "unsloth/gemma-4-12B-it-qat-GGUF", file: "mtp-gemma-4-12B-it.gguf"),
  extraArguments: ["-ngl", "999", "--spec-type", "draft-mtp"]
)

Bundled today

Profile	Quant	Draft	Download
Gemma 4 12B IT QAT	UD-Q4_K_XL	MTP	6.97 GB
Gemma 4 31B IT QAT	UD-Q4_K_XL	MTP	17.57 GB

Development

make run     # build, package, and launch the .app
make dist    # build the signed Dakodeon.app bundle
make zip     # build dist/Dakodeon.zip (release artifact)

Source layout

File	Responsibility
Catalog.swift	Curated model profiles + types
ModelStore.swift	Download / delete / status via the hf cache
ServerController.swift	llama-server lifecycle, active-model sync, shutdown
MenuView.swift	The menu bar panel
SettingsView.swift	Model-management window
DakodeonApp.swift	App entry, scenes, and icons

License

MIT for the app. Model weights remain under their own licenses — the bundled Gemma model follows the Gemma Terms of Use.