# Local models

> Run the agents on a model on your own computer, through Pi.

Pi can talk to any server that speaks an API it supports, including Ollama, LM Studio, llama.cpp and vLLM. Nothing
leaves your computer, and there is no key to buy.

## Add the server to Pi

Describe the server in Pi's `models.json`, in `~/.pi/agent` (or in a project `.pi` folder). For Ollama:

```json
{
  "providers": {
    "ollama": {
      "baseUrl": "http://localhost:11434/v1",
      "api": "openai-completions",
      "apiKey": "ollama",
      "models": [{ "id": "qwen2.5-coder:7b" }]
    }
  }
}
```

The `apiKey` is a placeholder that makes the model available to Pi; Ollama ignores it. For a server that needs a real
key, use `"$NAME"` to read it from an environment variable. Pi's own documentation lists every field (its
`models.md` and `models-configuration.md`, in the `@earendil-works/pi-coding-agent` package).

Then run `pi` once so its model list is refreshed, restart Solipsist, and choose the model in the
[agent panel](https://solipsist.sh/docs/reference/agent-panel/): it's listed as `<provider>/<model>`, here `ollama/qwen2.5-coder:7b`.

## What to expect

- **Quality.** A small model follows a short rewrite well and an open-ended one badly. If its suggestions are poor, try
  a larger model before changing anything else.
- **Speed.** A run waits for your computer. Lower `agent.concurrency` (three runs at a time by default) if the machine
  struggles with several at once.
- **Cost.** A local run reports $0, so the dollar [spending limits](https://solipsist.sh/docs/agents/how-agents-work/#spending-limits)
  never stop it.
- **Tools.** Agents give their result by calling a tool, so use a model that supports tool calls. If a model can't,
  its runs will fail: pick another.

If it doesn't work, see [When it goes wrong](https://solipsist.sh/docs/engines/when-it-goes-wrong/).

Source: https://solipsist.sh/docs/start/local-models/