hobhob

Your own model provider

Run hob on Ollama, LM Studio, OpenRouter, or another OpenAI-compatible server that you connect.

hob is the agent that comes with the app. Like any agent, it needs a model. A license can include hob inference, which gives hob its models. Your own model provider is the other way: you connect a model server, and hob uses its models.

You can connect:

  • Ollama or LM Studio, on this computer or on another one.
  • OpenRouter, for hosted models. It needs a key.
  • Another server that offers the OpenAI-compatible API, for example llama.cpp or vLLM.

Each provider you connect shows as its own agent, with the name you give it. You choose it in the agent menu as you choose any other agent.

Your provider bills you, not hob

OpenRouter bills your OpenRouter account for this use. Your hob plan does not include it. A server that you run yourself, such as Ollama or LM Studio, uses your own computer.

Connect a provider

Start in one of two places:

  • During setup, at Connect your agents: the Your own provider row, then Connect.
  • At any time: Settings → Agents, then + Provider.
  1. Choose the provider. hob first looks for Ollama and LM Studio on this computer and lists the servers it finds. Choose a server that it found, or choose the kind of server. A server that hob found has two choices: Add connects it at once, with all of its models that work with hob; Choose models lets you choose them first.
  2. Give the address and the key. hob fills in the usual address for Ollama (localhost:11434) and LM Studio (localhost:1234). For a server on another computer, use its name or network address, with the port. Give a key only if the server needs one; OpenRouter always does. Then select Test.
  3. Choose the models. hob tests each model and shows which ones work with hob, and why the others do not. Select the models to show in the agent menu, and give the provider the name that agent menus show. If the server does not report a model's context, you can type it; if you leave it empty, hob uses the model's maximum.
  4. Add it. The provider shows in your list of agents and as its own agent in the agent menu.

hob keeps the key encrypted in its secret store on the computer where hob runs, and sends it only to the provider's own address. If a server sends hob to another address, hob does not follow it on its own, so a key stays with the address you gave.

What a model needs to work with hob

hob does its work through tools, so a model must be able to use tools. A model that cannot use tools does not work with hob. vLLM offers tools only when it starts with --enable-auto-tool-choice and --tool-call-parser.

hob needs at least 16,384 tokens of context; 32,768 or more is better. With less, the server cuts the start of the conversation without a warning. To give a model more context:

ServerWhat to do
OllamaIn the Ollama app's Settings, move the Context length slider to 32,768 or more, or set OLLAMA_CONTEXT_LENGTH=32768. Quit Ollama and start it again.
LM StudioLoad the model again with a context length of 32,768 or more: lms load <model> --context-length 32768, or set it in the model's load settings.
llama.cppStart the server again with -c 32768 or more.
vLLMStart vLLM again with --max-model-len 32768 or more.

Then select Check again. hob tests the model again and shows it as usable.

When a test fails

hob tells you what went wrong and what to do next. These are the usual causes.

  • Nothing answers at the address. Start the server: open Ollama, or in LM Studio open the Developer tab and turn on Start server (or run lms server start). Then select Try again.
  • The server listens only to its own computer. Ollama and LM Studio answer only to the computer they run on until you change one setting there:
    • Ollama: set OLLAMA_HOST to 0.0.0.0:11434. On macOS, run launchctl setenv OLLAMA_HOST "0.0.0.0:11434", then quit and open the Ollama app. On Windows, set the OLLAMA_HOST environment variable for your user account and start Ollama again. On Linux, add Environment="OLLAMA_HOST=0.0.0.0:11434" to the ollama service and restart it.
    • LM Studio: in its server settings, turn on Serve on Local Network, or start the server with lms server start --bind 0.0.0.0.
  • Another kind of server answers. hob found a different server at this address. Use the kind that hob offers, or change the address.
  • Something answers, but it is not a model server. The port may belong to another program. The usual ports are Ollama 11434, LM Studio 1234, vLLM 8000, and llama.cpp 9931 (8080 in older builds).
  • The server needs a key, or does not accept the key. Paste the key in the key field. A key that is not accepted may be mistyped, expired, or deleted. For OpenRouter, you can make a key on OpenRouter.
  • The server has no chat models. Download a chat model that can use tools, then check again.

Change or remove a provider

Open Settings → Agents and select the provider. Its page has the same choices that you made when you connected it:

  • Change the address, or replace, add, or remove the key. hob tests the provider again when you save.
  • Select Refresh model list after you add or remove models on the server, and Test to check the provider again.
  • In Models in the agent menu, show or hide each model, make one the default, and set the context hob uses.
  • Turn off In menus to hide the provider from agent menus and keep it.
  • Select Remove… to remove it. Panes that use it stop sending. Their conversations stay, and you can continue one with a new agent that references it. hob deletes the key at once, and your models stay on the server. Hide instead keeps the provider out of the menus without removing it.

hob inference and your own provider

When your license includes hob inference, the list of agents shows hob with hob inference: included with your license, and hob is ready without a provider. Your own provider works with or without hob inference.

On this page