> ## Documentation Index
> Fetch the complete documentation index at: https://docs.replayfy.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Enable AI

> Turn on Storylines, signals, investigations, and the Ask assistant with your own model key. One env var reaches every provider.

The full agentic AI ships in the open-source build — you bring your own model
key. On a self-host the key lives on the **server** (in the environment), shared
by every workspace, and is **unmetered**: you pay your provider directly, nothing
is metered by Replayfy.

AI is optional and off until you give it a provider. Set three variables and
restart the API.

## Any provider, one key (recommended)

Point `LLM_PROVIDER` at **OpenRouter** and you reach every provider by model id —
no per-provider setup:

```bash theme={null}
LLM_PROVIDER=openrouter
OPENROUTER_API_KEY=sk-or-...
LLM_MODEL=anthropic/claude-sonnet-4.5
```

Swap `LLM_MODEL` for any model OpenRouter offers:

| Provider     | Example `LLM_MODEL`            |
| ------------ | ------------------------------ |
| Anthropic    | `anthropic/claude-sonnet-4.5`  |
| OpenAI       | `openai/gpt-4o`                |
| Google       | `google/gemini-2.5-pro`        |
| xAI          | `x-ai/grok-4`                  |
| DeepSeek     | `deepseek/deepseek-chat`       |
| Mistral      | `mistralai/mistral-large`      |
| Groq (Llama) | `groq/llama-3.3-70b-versatile` |

## Anthropic directly

If you'd rather call Anthropic without a gateway:

```bash theme={null}
LLM_PROVIDER=            # blank = Anthropic direct
ANTHROPIC_API_KEY=sk-ant-...
LLM_MODEL=claude-sonnet-4-5   # bare model id, no vendor prefix
```

## What it powers

<CardGroup cols={2}>
  <Card title="Storylines" icon="book-open">
    A narrated summary of what happened across your product.
  </Card>

  <Card title="Signals & investigations" icon="magnifying-glass">
    Surfaced issues, with a grounded root-cause pass on demand.
  </Card>

  <Card title="Ask" icon="message">
    Ask about a metric, funnel, or session in plain language.
  </Card>

  <Card title="Likely cause" icon="lightbulb">
    The Overview's "why did this move" explanations.
  </Card>
</CardGroup>

## No key set?

Nothing crashes. AI features tell you they need a key: the Ask panel shows a
"set up AI" notice instead of a broken chat, and **Settings → AI** reports
`Not configured` with the exact env vars to set. Add the key, restart, done.

<Note>
  The key is server-side by design — one key for every workspace, and it stays out
  of the application database. There is no per-workspace key entry in the
  self-hosted dashboard.
</Note>