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

# Attachments, size limits and media generation

You can give the agent images, PDFs, audio and video, and it can read such files in your project.
What works depends on the model. This page lists what each provider accepts, the size limits, and
how to let the agent generate images and video.

## Attach a file

| Where | How |
| --- | --- |
| Terminal UI | Drag a file into the terminal or paste its path; it becomes a chip |
| Desktop | **Attach files**, or drop files on the chat |
| One prompt | `haarness run -f photo.png -f spec.pdf "…"` |
| API and SDKs | `attachments` on the turn; the SDKs' `attachments()` helper reads local files |

The agent can also open media files in your project with its `read` tool. They reach the model the
same way as an attachment.

## What each provider accepts

| Provider | Images | PDF | Audio | Video |
| --- | --- | --- | --- | --- |
| OpenAI | Yes | Yes | No | No |
| Anthropic | Yes | Yes | No | No |
| Amazon Bedrock | Yes | Yes | No | No |
| Gemini and Vertex AI | Yes | Yes | Yes | Yes |
| Azure, Copilot and OpenAI-compatible | If the model supports them | Yes | MP3, WAV | No |

The model has the last word: an attachment it cannot use returns an error that says so. For audio
and video, use a Gemini model.

For a [custom or local provider](https://haarness.ai/docs/providers/custom-providers.md), images are off until you set
`"images": true`, because Haarness cannot tell whether the model accepts them.

## Size limits

| Limit | Applies to |
| --- | --- |
| 20 MiB in total | The attachments on one message |
| 10 MiB in total | The attachments on one message to Gemini or Vertex AI |
| 20 MiB | One media file the agent reads; see [execution budgets](https://haarness.ai/docs/security/execution-budgets.md) |
| 10 MiB | One file opened or saved in Desktop's file explorer |
| 50 MiB | One generated image or video |

Desktop also accepts at most 20 files per message.

When you are attached to a server on another machine, the files you attach are uploaded with your
message, so the server does not need access to your disk.

## When a chat gets too long

Every model has a context window. When a chat fills most of it, Haarness summarises the
older part and keeps the recent turns word for word, so you can carry on. Type `/compact` in the
terminal UI to do it now.

You can turn this off or change when it happens with the `compaction` settings in
[Configuration](https://haarness.ai/docs/cli/configuration.md). Haarness takes each model's context and output size from
the provider, or from its own catalogue when the provider does not say. For a custom provider, set
`context_window` and `max_output_tokens` yourself.

## Generate images and video

The agent can create an image or a video when you ask for one. Generation is off until you
configure a backend:

| Tool | Turn it on with |
| --- | --- |
| Images | `OPENAI_API_KEY` in the environment (uses OpenAI's image API), or `HAARNESS_IMAGE_GENERATE_URL` |
| Video | `HAARNESS_VIDEO_GENERATE_URL` |

A stored OpenAI key or a ChatGPT sign-in does not turn on image generation; the environment
variable does. Set it before Haarness starts its server.

Each generation asks for your permission first, because it costs money and writes a file, unless
your [permission rules](https://haarness.ai/docs/security/permissions.md) already allow it (`"media_generate": "allow"`).
In a chat in Auto, Auto's reviewer answers that request instead of you.
The result is saved in `.haarness-chat-attachments/` in your project, which Haarness keeps out
of Git, and shown in the chat.

### Use your own generation service

Point `HAARNESS_IMAGE_GENERATE_URL` or `HAARNESS_VIDEO_GENERATE_URL` at an HTTPS endpoint
(`http://localhost` also works). Haarness sends a JSON `POST` with `model` and `prompt`, plus
`size` for images or `aspect_ratio` and `duration` for video. If `HAARNESS_MEDIA_API_KEY` is set,
it goes in an `Authorization: Bearer` header; your OpenAI key is never sent there.

Reply with JSON containing the base64 file in `b64_json`, `base64` or `data`, and optionally its
type in `mime_type`. Haarness checks that the bytes are really an image or video.

## Related

- [Choose a provider and model](https://haarness.ai/docs/providers/overview.md)
- [Permissions](https://haarness.ai/docs/security/permissions.md)