Documentation

OpenHydra wraps the AI engine you already run and exposes one OpenAI-compatible endpoint at http://127.0.0.1:16527/v1. Point any tool at it, locally, or across the peer-to-peer network when you need a bigger model.

OpenHydra runs AI models right on your own computer, and connects you to other people’s computers when you want something bigger. Here’s how to get going, all in the app, no coding required.

Getting started

Three steps, no account, no config:

Getting started takes about a minute, no account, and nothing to set up:

1. Install the app for macOS, Windows or Linux (the macOS build is signed & notarized).
2. Install the app for Mac, Windows or Linux, just like any other program. On Mac it opens straight away, no scary warnings.
3. Run a model. OpenHydra detects an engine already on your machine, or installs one for you (see Engines). It serves that model locally at the endpoint below.
4. Pick a model and press go. A “model” is the AI brain that answers you. Don’t have one yet? OpenHydra installs one for you with a click, nothing to set up by hand.
5. Point your tools at http://127.0.0.1:16527/v1 with the model id openhydra/auto. Flip to network mode any time to reach models peers are serving.
6. Start chatting. Type in the app and the AI replies, running right on your computer. Want a bigger, smarter one? Flip a switch to borrow power from the network.

You usually won’t do this by hand, you just chat in the app, and it sends your messages to the AI for you. “Sending a request” simply means an app asking the AI something and getting an answer back. That same ability is what lets other apps use OpenHydra too, that’s what Connectors are for.

Your first request

The endpoint speaks the OpenAI API, so a plain curl works:

curl http://127.0.0.1:16527/v1/chat/completions \
  -H "Authorization: Bearer oh-local" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openhydra/auto",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": true
  }'

Any non-empty API key is accepted locally. openhydra/auto lets the router pick a model; you can also name a specific one a provider announces. /v1/chat/completions (with SSE streaming), /v1/embeddings and /v1/models are all supported.

Engine discovery & install

OpenHydra doesn't ship an inference engine, it discovers the ones already running on your machine and can install the ones you don't have yet. Supported engines:

An engine is the program that actually runs an AI model on your computer, think of the model as the “brain,” and the engine as the body it runs in. OpenHydra doesn’t include one; it looks for an engine you already have, or installs one for you with a single click. You never need to touch the technical bits below, they’re only there if you enjoy tinkering. The engines it works with:

• Ollama, general-purpose local LLMs (default).
• LM Studio, MLX-optimised models on Apple Silicon, with a local server.
• llama.cpp, lightweight GGUF runtime.
• vLLM, high-throughput serving (NVIDIA on Linux; Apple Silicon via the vllm-metal MLX plugin).
• Exo, shard a model across several of your own devices.
• ComfyUI, image generation (Stable Diffusion, Flux).

Discovery

On launch, OpenHydra probes your machine for a running engine on its standard port (Ollama :11434, OpenAI-compatible :8000, LM Studio :1234, llama.cpp :8080, ComfyUI :8188) and detects installed engines wherever they live, a CLI on your PATH (including Homebrew), a /Applications app, or a known virtual environment. If one is already serving, OpenHydra uses it as-is; you don't have to point at anything.

When you open OpenHydra, it automatically looks around your computer to see if you already have an engine running, and just uses it if so. You don’t have to point it at anything or copy any settings, it works this out on its own.

One-click install

From Engines in the app, missing engines install with a consent → live progress → auto-detect flow. Downloads OpenHydra performs itself show a determinate bar (speed & ETA); vendor installers stream their own progress. What runs is chosen per operating system and verified against each vendor's docs:

Don’t have an engine yet? In the app’s Engines screen, click Install next to one and OpenHydra downloads and sets it up for you, showing progress as it goes. When it’s finished, it’s ready to use. The table below shows exactly what gets installed on each kind of computer, handy if you’re curious, but you never have to run any of it by hand.

After install, an engine card is one of three states: Install → not installed · Run → installed & idle (self-serving engines) · Manage → running.

Connectors

OpenHydra is an OpenAI-compatible endpoint at http://127.0.0.1:16527/v1. Point any tool with a custom base URL at it and use the model id openhydra/auto (or any model your provider announces). direct plugs in as-is · shim needs an Anthropic→OpenAI translating proxy (e.g. LiteLLM) in front.

A connector lets you use OpenHydra’s AI inside another app, like a coding assistant, your code editor, or any tool with an AI feature, instead of only chatting in the OpenHydra app. Here’s the trick: most of these apps were built to talk to ChatGPT, and OpenHydra speaks that exact same language. So in the app’s settings you just paste one web address, http://127.0.0.1:16527/v1, pick your model, and the app quietly starts using OpenHydra instead. Same app you know, but now the AI is free and running on your own computer. Ready-made settings for popular tools are below, find yours and copy them in.

export OPENAI_BASE_URL=http://127.0.0.1:16527/v1
export OPENAI_API_KEY=oh-local
# model: openhydra/auto

provider: openai
apiBase: http://127.0.0.1:16527/v1
model: openhydra/auto

# Anthropic API → run a translating proxy
# (LiteLLM) in front of OpenHydra
ANTHROPIC_BASE_URL=http://127.0.0.1:4000

openclaw config set model.baseUrl \
  http://127.0.0.1:16527/v1
openclaw config set model.name openhydra/auto

adapter = "openai_compatible"
url = "http://127.0.0.1:16527/v1"
model = "openhydra/auto"

(gptel-make-openai "OpenHydra"
  :host "127.0.0.1:16527" :protocol "http"
  :models '("openhydra/auto"))

pi config set api.baseUrl \
  http://127.0.0.1:16527/v1
pi config set api.model openhydra/auto

# Settings → Model provider → OpenAI-compatible
base_url: http://127.0.0.1:16527/v1
model: openhydra/auto

Sharing your machine

Running an engine already? Turn on Share to serve the herd while your machine is idle, you earn priority for when you need a model bigger than your own, plus reputation that keeps the network honest. Routing, encryption and NAT traversal are handled for you; you choose which models to offer and can stop with one toggle.

Got a computer that’s often switched on but not doing much? You can let others use its spare power to run AI, and in return, you get to use the network’s bigger models when you need them. Just switch on Share and pick which models you’re happy to offer. Staying private, connecting safely, and everything else technical is handled for you, and you can stop any time with one switch.

FAQ