Can AI agents run on my own models?
Yes. An agent can use a model on your own computer through Ollama, LM Studio or vLLM, as long as the agent software accepts an OpenAI-style endpoint. It costs nothing per message and nothing leaves your network. The trade-off is skill on long, tool-heavy tasks.
What "your own model" can mean
| Option | What it is | You pay |
|---|---|---|
| A local model | A model running on your machine or network through Ollama, LM Studio or vLLM | Hardware and electricity |
| Your own API key | You buy model usage straight from a provider | The provider's price, with no middleman |
| A plan you already have | Sign in with a subscription such as ChatGPT instead of a key | Your existing subscription |
All three need an agent that lets you choose the model. Many hosted agents don't: they run a fixed model and bill you for it.
What local models are good at
- Privacy. Prompts, files and results stay on your machine.
- Cost. No per-message charge, so routine work is free.
- Simple, frequent jobs: sorting, summarising, tagging, drafting short replies.
- Embeddings for memory and search.
Where they fall short
- Long tasks with many tools. Agents call tools dozens of times in a row. Smaller models lose track sooner, pick the wrong tool more often and recover from errors less well.
- Browsing and screenshots. Reading a web page from an image needs a model with vision.
- Speed. Without a strong GPU, each step is slow, and a task has many steps.
A setup that works
Most people who run local models don't use them for everything:
- A local model for private or simple bots.
- A hosted model, on your own key, for hard tasks.
- A fallback: if the first model is unavailable, the bot uses the second.
How to connect one
Start the model server, then give your agent its address. For Ollama on the same machine that is usually http://127.0.0.1:11434/v1. Pull a model with tool support first, and a separate small model for embeddings if the agent uses memory search.
Local models in Casama
Self-hosted Casama accepts any OpenAI-style server: Ollama, LM Studio, vLLM or your own. Each bot has a main model and fallbacks, so one bot can run locally while another uses a provider key. Calls to a local model are recorded in Usage at $0. You can also use keys for Anthropic, OpenAI, Google Gemini, xAI or OpenRouter, or sign in with a ChatGPT plan.
On Casama Cloud, Casama picks and runs the models for you. Choosing your own model is part of self-hosting, which is free.
Questions
Which local model should I use?
Pick the largest model with tool calling that your hardware runs at a usable speed, and test it on a real task. The bots and models guide has current advice.
Does the bot's computer need a GPU?
No. The model runs wherever your model server is. The bot's computer only runs the tools.
Can I mix local and hosted models?
Yes, per bot, with fallbacks.