gitwtfhub

wtf is laguna-otari-bridge?

njbrake/laguna-otari-bridge — explained in plain English

Analysis updated 2026-05-18

2ShellAudience · developerComplexity · 5/5LicenseSetup · hard

TL;DR

A set of scripts that run poolside's large Laguna AI model locally on an Apple Silicon Mac and expose it as a password-protected, internet-accessible API.

Mindmap

mindmap
  root((Laguna Bridge))
    What it does
      Runs Laguna model locally
      Exposes it as an API
      Secures access with a gate
    Tech stack
      llama.cpp
      Caddy
      Tailscale
    Use cases
      Local AI instead of cloud
      Secure remote API access
      Avoid known setup pitfalls
    Audience
      Developers
      Mac power users
      Self hosters

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

Why would anyone build with this?

REASON 1

Run a large open-weight AI model locally on a Mac instead of paying for a cloud API.

REASON 2

Expose a locally running AI model as a secure, internet-reachable API endpoint for a client app.

REASON 3

Learn practical gotchas around running a specific large model with llama.cpp on Apple Silicon before hitting them yourself.

REASON 4

Set up an authenticated bridge in front of a local model server so only requests with the right key get through.

What's in the stack?

Shellllama.cppCaddyTailscale

How it stacks up

njbrake/laguna-otari-bridgeaizhigitovamir-code/n8n-automation-suiteayush-sk-pathak/claude-code-starter-pack
Stars222
LanguageShellShellShell
Setup difficultyhardmoderateeasy
Complexity5/53/52/5
Audiencedeveloperpm founderdeveloper

Figures from each repo's GitHub metadata at analysis time.

How do you spin it up?

Difficulty · hard Time to first run · 1day+

Requires a Mac with at least 128GB of unified memory, building a custom llama.cpp fork, and downloading a 75GB+ model file.

The scripts in this repo are MIT licensed, free to use and modify, the Laguna model weights themselves use poolside's own separate license.

Wtf does this do

Laguna Otari Bridge is a set of scripts for running a large AI language model called Laguna, made by a company called poolside, directly on an Apple computer with an M series chip, and then making that model available over the internet as a private, password protected API endpoint. The project explains how to build a special version of a tool called llama.cpp that can run the Laguna model using the Mac's graphics hardware, download the model's weights, which are the large files that make up the model itself, and start a local server that answers requests in the same format used by OpenAI's API. A second part of the project, called otari bridge, puts a security gate in front of that local server using a tool called Caddy and a service called Tailscale, so the model can be reached from outside the home network over a secure web address while still requiring an access key. Most of the README documents things the author discovered while getting this working, since much of it was not written down anywhere else. For example, an optional speed up feature recommended for the model actually made responses several times slower here, because none of its predictions were ever accepted, so the author suggests leaving that feature off until the underlying cause is understood. The README also notes that the model's own reasoning text uses up part of the same token budget as its final answer, so a low limit on response length can cause an empty looking reply, and that the underlying server will keep answering with whichever model it loaded regardless of what a client asks for by name, which the security gate in this project corrects. Running this requires a Mac with Apple Silicon and a large amount of unified memory, at least 128 gigabytes for the smaller of the supported model file sizes, along with roughly 80 gigabytes of free disk space and, optionally, a Tailscale account for the public facing option. The scripts in this repository are released under the MIT license, so they can be freely used, changed, and shared, including commercially. The Laguna model itself is released under a separate license set by poolside, not by this project.

Yoink these prompts

Prompt 1
Help me build poolside's llama.cpp fork with Metal support to run the Laguna model on my Mac.
Prompt 2
Walk me through downloading the Laguna model weights and starting llama-server with the right context and GPU settings.
Prompt 3
Explain why speculative decoding is currently making Laguna slower instead of faster on Metal.
Prompt 4
Show me how to set up the otari-bridge Caddy and Tailscale gate to expose my local model securely.

Frequently asked questions

wtf is laguna-otari-bridge?

A set of scripts that run poolside's large Laguna AI model locally on an Apple Silicon Mac and expose it as a password-protected, internet-accessible API.

What language is laguna-otari-bridge written in?

Mainly Shell. The stack also includes Shell, llama.cpp, Caddy.

What license does laguna-otari-bridge use?

The scripts in this repo are MIT licensed, free to use and modify, the Laguna model weights themselves use poolside's own separate license.

How hard is laguna-otari-bridge to set up?

Setup difficulty is rated hard, with roughly 1day+ to a first successful run.

Who is laguna-otari-bridge for?

Mainly developer.

View the repo → Decode another repo

This repo across BitVibe Labs

Don't trust strangers blindly. Verify against the repo.