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.
💤  Idle model sleep	The router stays online, loads models on demand, and unloads model memory after inactivity.
🧹  Clean shutdown	Quitting the app stops llama-server.
🚀  Native defaults	llama-server loads each model's trained context and embedded chat template.
🧩  Curated assets	Profiles can include weights, draft models, vision projectors, and tuned llama-server flags.

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 gemma4-26b-a4b-it-qat and ornith-1.0-35b. 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.

The router process is designed to stay up for long sessions. Dakodeon starts it with on-demand model loading and a five-minute idle sleep window: the model process and KV cache are released after inactivity, while the OpenAI-compatible endpoint remains available and reloads the requested model on the next task. The menu and Settings window show whether the active model is loaded, loading, sleeping, or unloaded.

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, an optional vision projector, an optional chat template, and any extra llama-server flags. The app exposes no per-user configuration — to add a model, append an entry:

ModelProfile(
  id: "ornith-1.0-35b",
  weights: ModelAsset(repo: "deepreinforce-ai/Ornith-1.0-35B-GGUF", file: "ornith-1.0-35b-Q4_K_M.gguf"),
  draft: nil,
  mmproj: nil,
  chatTemplate: ModelTemplates.ornith,
  extraArguments: ["-ngl", "999", "--reasoning-format", "deepseek"]
)

Bundled today

Profile	Quant	Draft	Vision	Download
Gemma 4 26B A4B IT QAT	UD-Q4_K_XL	MTP	✓	15.69 GB
Ornith 1.0 35B	Q4_K_M	-	-	21.17 GB

Ornith uses its Hugging Face chat template with a small Codex compatibility patch: developer messages are rendered as system messages, and system messages are not required to appear only at index zero. Thinking remains enabled; llama-server is started with --reasoning-format deepseek so <think> content is returned as structured reasoning instead of being printed in the assistant answer.

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 profile follows the Gemma Terms of Use.