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

# Local deployment

> Run everything on one machine with Ollama.

The default deployment mode. Ollama handles inference, Bumblebee handles everything else — all on your machine.

## Prerequisites

<Check>Ollama installed and in your PATH</Check>
<Check>Models pulled: `gemma4:26b` and `nomic-embed-text`</Check>
<Check>GPU with 16 GB+ VRAM (see [Hardware](/reference/hardware))</Check>

## Running

<Tabs>
  <Tab title="CLI only">
    ```bash theme={null}
    bumblebee talk canary --ollama
    ```

    Direct conversation, no daemon. Good for testing and development.
  </Tab>

  <Tab title="Full daemon">
    ```bash theme={null}
    bumblebee run canary --ollama
    ```

    Heartbeat, soma, memory consolidation, initiative, and every platform in entity YAML.
  </Tab>

  <Tab title="Fresh machine">
    ```bash theme={null}
    bumblebee run canary --ollama --pull-models
    ```

    Also downloads chat + embedding models from the Ollama library.
  </Tab>
</Tabs>

The `--ollama` flag probes the configured URL (default `http://localhost:11434`) and runs `ollama serve` in the background if the endpoint is unreachable. It does not re-download models unless `--pull-models` is added.

## What runs

| Component     | Details                                                                 |
| ------------- | ----------------------------------------------------------------------- |
| **Ollama**    | Chat models (reflex + deliberate) and embedding model, loaded on demand |
| **Daemon**    | Heartbeat, soma ticks, memory consolidation, drive checks, wake cycles  |
| **Platforms** | Every platform in entity YAML (CLI, Telegram, Discord)                  |
| **Tools**     | Shell, filesystem, code execution — all on local machine                |

## Stopping

```bash theme={null}
bumblebee stop
```

| Flag                     | Effect                     |
| ------------------------ | -------------------------- |
| `--dry-run`              | Show what would be stopped |
| `--skip-gateway`         | Don't run gateway shutdown |
| `--leave-ollama-running` | Keep Ollama alive          |

## Ollama troubleshooting

<Warning>If Ollama falls back to CPU unexpectedly, run `npm run ollama:reset`. This stops everything, clears stale processes, sets safe defaults (`OLLAMA_MAX_LOADED_MODELS=1`, `OLLAMA_KEEP_ALIVE=60s`), restarts the gateway, and warms the model.</Warning>

## Configuration

```yaml theme={null}
deployment:
  mode: local

ollama:
  base_url: "http://localhost:11434"
  timeout: 120
  retry_attempts: 3

models:
  reflex: "gemma4:26b"
  deliberate: "gemma4:26b"
  embedding: "nomic-embed-text"
```

Override models per entity under `cognition.reflex_model` / `cognition.deliberate_model`.

<Note>
  **Optional:** To evaluate the same harness against **hosted frontier models** (OpenRouter, Venice) on this machine, switch `BUMBLEBEE_INFERENCE_PROVIDER` and set the provider key — local-first defaults and licensing are unchanged. See **[Hosted inference (testing)](/deployment/hosted-inference-testing)**.
</Note>