One endpoint, two dialects
A single local HTTP server speaks both the OpenAI and Anthropic APIs. Point any existing SDK or tool at localhost — no code changes.
Local + cloud LLM gateway
Stop playing with AI
and start engineering it.
Lattis is a desktop manager and gateway for LLMs. One local endpoint speaks both the OpenAI and Anthropic APIs, routing every request to local or cloud models — with usage and cost you can actually account for.
Open source · macOS, Linux, Windows · Rust + iced
~/lattis
$ curl localhost:1234/v1/chat/completions \
-d '"model": "qwen3-4b-instruct-2507", ...' # local GGUF
$ curl localhost:1234/v1/messages \
-d '"model": "claude-opus-4-8", ...' # cloud, OAuth
$ curl localhost:1234/v1/chat/completions \
-d '"model": "gpt-5.5", ...' # cloud, API key
→ one endpoint · OpenAI + Anthropic APIs · local + cloudWhat it does
A gateway, not another playground.
Local models, resident
Download GGUF models from Hugging Face and serve them with a bundled llama-server in router mode. Keep several resident; swap on demand.
Cloud, when you want it
Connect Anthropic and OpenAI / Codex via OAuth or API key. Requests are translated between OpenAI, Anthropic, and Responses formats automatically.
MLX on Apple Silicon
Detects Apple's mlx_lm if installed and serves MLX models alongside the rest — Metal-accelerated, hidden when unavailable.
Usage you can account for
Per-app, per-model, and per-project token usage with estimated cost. Know exactly what each project and provider is spending.
Runs as a daemon
The daemon owns storage, downloads, the inference engine, and cloud connections. Launch on login and it serves the API 24/7 — app optional.
How it routes
One hub. Every model.
Behind a single endpoint, Lattis routes each request to the right engine — a resident local GGUF, an MLX model on Apple Silicon, or a connected cloud provider — translating API formats as needed so your client never has to care.
See the API →Get going
Three steps to a local AI backbone.
- 01
Install
Grab the desktop app (coming soon) or build from source.
- 02
Add a model
Download a GGUF from the Library, or connect a cloud provider.
- 03
Point & ship
Aim your client at localhost:1234 and build.