Puente

Your own AI platform. Runs on your hardware.

Puente is a local-first AI orchestrator. It stands up a full stack of AI services — chat, image generation, speech, search, notebooks — on modest hardware, from one config file. Private by default. Yours to keep.

Pay a time tax, not a token tax. Modest hardware, honest bills.

Puente is the deployment layer of LocoPuente, a LocoLabo initiative: local AI for everyone who can't — or won't — send their data to the cloud.

Install

pip install locopuente

The PyPI distribution is locopuente (the puente name belongs to another project). The command you run is puente.

Or from source:

git clone https://github.com/michael-borck/loco-puente.git
cd loco-puente && pip install -e .

Quick start

puente init      # detect your hardware, choose services that fit
puente install   # pull Docker images, install native pieces (Ollama, models)
puente up        # start the stack + a launcher portal

Then puente status to see what's live. That's it.

Expose services (optional)

Enable the built-in Caddy service and give any service a proxy: block to publish it at a public hostname — automatic TLS, per-service auth. It's config-as-code: you declare the boundary, Caddy serves it.

services:
  caddy:
    enabled: true
    email: you@example.org         # Let's Encrypt contact
    users:
      ui: { admin: ADMIN_BCRYPT }  # basic-auth groups (bcrypt hash from env)
    proxy_hosts:                   # front hosts that aren't puente services, too
      - host: plex.example.org
        port: 32400
        upstream: 192.168.1.10
        auth: none

  swarmui:
    proxy:
      host: swarmui.example.org
      auth: bearer                 # none | basic | bearer
      token_env: SWARMUI_TOKEN     # token read from the environment

Secrets (ADMIN_BCRYPT, SWARMUI_TOKEN, …) come from the environment, never the committed config. Full walkthrough — including migrating off an existing proxy — in docs/caddy-migration.md.

What it does

• Detects your hardware and proposes a service set it can actually run, pinning models to the right GPU.
• Orchestrates the containers — you toggle services in puente.yml, Puente handles Docker, GPUs, models, and (optionally) a reverse proxy.
• Config as code. The whole stack is one committable puente.yml. No web UI, no hidden state.
• Coexists with what you already run. A service you've already installed can stay (managed: false); Puente uses it instead of spinning up its own.

Services

A pick-and-choose menu, all running locally:

Service	What it is
Ollama	Local LLM inference
Open WebUI	Chat over your models
SwarmUI / ComfyUI	Image generation
Voicebox / Speaches	Voice-cloning TTS, speech-to-text
SearXNG	Private meta-search
AnythingLLM	Docs + RAG workspaces
Open Notebook, Stirling PDF, Excalidraw, Jupyter, …	Tools

Plus an optional Caddy reverse-proxy service (automatic TLS) that fronts whichever services you expose — see docs/caddy-migration.md.

Most services pull a prebuilt image and start in seconds. Voicebox is the exception: upstream publishes a Dockerfile but no image, so the first puente up voicebox compiles it from source (torch + CUDA — budget 10–20 minutes, with long silent stretches). It's a one-off; later starts reuse the built image.

Commands

puente init      Interactive setup — detect hardware, pick services
puente install   Install native services + pull Docker images
puente up        Start the stack (or a specific service)
puente down      Stop the stack (or a specific service)
puente enable    Enable a service in puente.yml
puente disable   Disable a service in puente.yml
puente status    Status of all enabled services
puente doctor    Health-check enabled services
puente gpu       Detect and display GPUs
puente connect   Connection details for external tools
puente portal    Generate the service launcher page

Requirements

• Python 3.10+
• Docker (for containerized services)
• A GPU is recommended but not required — a single consumer GPU is enough to start.

License

MIT © Michael Borck. A LocoLabo initiative, Curtin University.