Image generation

Ask Flowly for an image and it appears inline in the reply — no separate message, no link to click. Generation runs through FAL with a model you choose, the file is downloaded to your Flowly home, and Flowly's media-delivery layer surfaces it on whatever surface you're talking from: a Telegram photo, an inline image in the desktop/iOS chat, or an attachment fetched over the gateway.

It's an opt-in tool: it registers only when tools.image_generation is enabled and has an API key. Bring your own FAL key — there's no hosted image quota in the open-source build.

Quick start

> draw a watercolor fox sleeping under a maple tree, autumn light
… Flowly: Here's the watercolor fox — soft autumn palette, curled up under
  the maple. (image attached)

The agent decides to call image_generate on its own when your message asks for a picture. You never invoke the tool directly.

Setup

The tool is gated behind the FAL Image Generation integration card. Configure it any of three ways:

1. Setup wizard (recommended)

flowly setup            # full wizard — pick "FAL Image Generation", paste a key, choose a model

Get a key at fal.ai/dashboard/keys.

2. Desktop / remote — open Connections, find FAL Image Generation, enter your key and pick a model. This works for local, relay, and self-hosted gateway bots through the same connection RPC.

3. By hand — edit ~/.flowly/config.json (keys are camelCase):

{
  "tools": {
    "imageGeneration": {
      "enabled": true,
      "apiKey": "fal-...",
      "model": "fal-ai/flux/dev"
    }
  }
}

Restart the gateway / start a new session after enabling — the tool is wired at agent boot.

Models

FAL has no clean "list all image models" API, so Flowly ships a curated short list of popular, known-good models. You pick the default in setup; the agent can override it per call via the model parameter.

You can also pass any FAL text-to-image model id as the model argument — the curated list is just what the picker shows.

The image_generate tool

How delivery works

The generated file is downloaded to ~/.flowly/media/img-<id>.png and then rides the assistant's own reply — there is no separate "here's your image" message. The same picture reaches every surface through Flowly's existing media path:

• Messaging channels (Telegram, WhatsApp, …) — sent as a native photo/attachment.
• Relay-connected apps (iOS / desktop / web) — uploaded and surfaced from the conversation.
• Direct gateway clients (self-hosted iOS / desktop) — delivered as an inline thumbnail with the reply, with the full-resolution original served on demand from GET /api/media?id=… (tap to zoom). No relay/S3 needed.

Because the file lives on disk, a generated image still shows correctly when you re-open a past conversation.

Configuration reference

Everything lives under tools.imageGeneration in ~/.flowly/config.json:

Notes & limits

• Bring your own key — image generation bills to your FAL account; there is no hosted quota in the open-source build.
• Synchronous — Flowly calls FAL's sync endpoint (POST https://fal.run/{model}), which suits images (a few seconds). Errors (auth, network, malformed response) come back as a short, user-facing message.
• num_images is capped at 4 per call.
• Image-editing (image-to-image) is not wired yet — the model catalog records an edit_endpoint for FLUX [dev] for a future image_edit tool, but there's no edit tool in this version.

Related

• Tools reference
• Voice · Artifacts
• Feature overview