---
title: "Free API access"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Free API access}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
knitr::opts_chunk$set(eval = FALSE)
```


*You can use codriver without paying for an API subscription!*

Codriver requires API access to an LLM provider, but isn't tied to a particular
provider or subscription. There are several free options available: **free API
access** with a cloud provider, or setting up your own local server. This
article covers some available options that are free to start with, and some
considerations of each.

------------------------------------------------------------------------

## Google Gemini

Google offers free API access to Gemini models through
[Google AI Studio](https://aistudio.google.com). Sign in with a Google account,
create an API key, and store it in your `.Renviron`:
```r
file.edit("~/.Renviron")
```

Add:
```
GOOGLE_API_KEY=*your_key_here*
```

Save the file, *restart RStudio* and then configure codriver:
```r
codriver::codriver_configure("google_gemini", model = "gemini-3.5-flash-lite")
```

At the time of writing **Gemini 3.1 Flash Lite** or **Gemini 3.5 Flash Lite**
are the recommended model for codriver on the free tier. These models offers 500
requests per day each (max. 15 per minute) and respond reasonably quick. Other
Gemini models on the free tier are either heavily rate-limited or slow enough to
make interactive use impractical.

> To check current free tier limits, go to
[Google AI Studio](https://aistudio.google.com) > *Usage & billing* > *Rate
limits*. Sort by *Category* to find *text-out models*.

On the free tier, Google may use your prompts to improve their models.
Review [Google's API terms](https://ai.google.dev/gemini-api/terms) before use.

------------------------------------------------------------------------

## Mistral

Mistral AI, a French company, offers a free tier through
[console.mistral.ai](https://console.mistral.ai). Sign up, create an API key,
and store it in your `.Renviron`:
```r
file.edit("~/.Renviron")
```

Add:
```
MISTRAL_API_KEY=*your_key_here*
```

Save the file, *restart RStudio* and then configure codriver:
```r
codriver::codriver_configure("mistral", model = "mistral-medium-latest")
```

At the time of writing, **Mistral Medium Latest** is the recommended model for
codriver on the free tier, but you may run into rate limits. Start with
`mistral-medium-latest`, and if you run into rate limits, switch to the
highest-numbered `mistral-medium-*` version available in the model list:

```r
# Get available models and filter on `mistral-medium-*`
ellmer::models_mistral() |>
  dplyr::filter(startsWith(name, "mistral-medium-"))

# Look up highest-numbered `name` and replace `2508` below
codriver::codriver_configure("mistral", model = "mistral-medium-2508")
```

The free tier of Mistral comes with the same caveat as Google Gemini: your
prompts may be used for model training. That said, Mistral is a European company
subject to EU law, which may be relevant for data governance considerations.
Check [Mistral's terms of service](https://mistral.ai/terms) for the current
policy.

------------------------------------------------------------------------

## Local model server

For maximum privacy, run a model locally. [Ollama](https://ollama.com) and
[LM Studio](https://lmstudio.ai) both provide local model servers that codriver
can connect to. No API key required, and nothing leaves your machine.

The trade-off is speed and model capability. Local models are generally slower
and less capable than cloud models, especially on machines without a dedicated
GPU. For lightweight tasks like completing a short line of code they can work
well; for more complex generation or editing, results may be less reliable.

##### Ollama

After installing and pulling a model, e.g. `ollama pull codellama`, configure
codriver:
```r
codriver::codriver_configure("ollama", model = "codellama")
```

##### LM Studio

Start your local server in the [LM Studio](https://lmstudio.ai) app, then
configure codriver:
```r
codriver::codriver_configure("lmstudio", model = "your-loaded-model")
```

------------------------------------------------------------------------

## Other options

The free tier landscape changes quickly. A few other providers worth knowing
about:

- **[OpenRouter](https://openrouter.ai)**

  A gateway to many models from different
  providers, some of which are free. One API key gives access to a wide range of
  models. Use `"openrouter"` as the provider name in codriver.

- **[Groq](https://console.groq.com)**

  Fast inference on open models like Llama and Mixtral. Free tier available. Use
  `"groq"` as the provider name in codriver.