higgsfield

name: higgsfield description: Generate images, videos, edits, or speech using Higgsfield AI via browser automation disable-model-invocation: true allowed-tools: - mcp__claude-in-chrome__navigate - mcp__claude-in-chrome__computer - mcp__claude-in-chrome__find - mcp__claude-in-chrome__read_page - mcp__claude-in-chrome__form_input - mcp__claude-in-chrome__get_page_text - mcp__claude-in-chrome__take_screenshot - mcp__claude-in-chrome__javascript_tool - mcp__claude-in-chrome__tabs_context_mcp - mcp__claude-in-chrome__tabs_create_mcp - mcp__claude-in-chrome__upload_image - Read triggers: - "higgsfield" - "higgsfield generate" - "generate with higgsfield" - "create image with higgsfield" - "create video with higgsfield" - "generate image" - "generate video" - "create image" - "create video"

GENERATE - Asset Generation via Higgsfield AI (Browser Automation)

Generate images, videos, edits, or speech by automating the Higgsfield website via Chrome browser tools. Uses the user's UNLIMITED website subscription instead of API credits.

Prerequisites

• Chrome must be open with the claude-in-chrome extension active
• User must be logged into higgsfield.ai
• If the user is not logged in, stop and tell them: "Please log in to higgsfield.ai in Chrome first, then retry."

Arguments

$ARGUMENTS describes what to generate. Examples:

• a cyberpunk city at night with neon signs
• video of a cat walking on a beach
• seedream photorealistic portrait of a woman in golden hour
• speech saying "hello world"
• edit this image to add sunglasses
• cinema studio cinematic shot of a forest

If no arguments provided, ask the user what they want to generate.

Model Selection Guide

When the user doesn't specify a model, pick the best one based on the task. Use the Best Model Recommendations table for quick selection.

Best Model Recommendations

Image Models (16 models)

Video Models (15+ models)

Editing Tools

Speech & Avatar

Special Features

Prompting Guide

Per-model prompting tips for best results:

• General: Short direct sentences, active verbs, one action per shot. Be specific about composition, lighting, style, mood.
• Nano Banana Pro: Use the 6-variable framework: Subject, Composition, Action, Location, Style, Technical. Example: "A samurai warrior, centered full-body shot, drawing katana, on a misty mountain bridge, ukiyo-e watercolor style, 4K detailed"
• Soul: Leverage the 60+ built-in style presets via the Style dropdown. Use Soul ID for character consistency across generations.
• Seedream 4.5: Direct like a photo shoot — specify lighting (golden hour, studio softbox), fabric details, soft focus areas. Best for fashion and portrait work.
• GPT Image 1.5: Use structured reasoning prompts with explicit layout instructions. Great for diagrams, infographics, and text-heavy compositions.
• Reve: Emphasize artistic style keywords — "concept art", "anime cel-shading", "oil painting". Supports up to 4 reference images.
• Z-Image: Keep prompts simple and direct for fastest results. Best for rapid ideation and iteration.
• Speak 2.0: Use stage directions in brackets [whispers], ALL CAPS for emphasis, ... for pauses. Example: "Welcome to... [dramatic pause] the FUTURE of AI."
• Cinema Studio: Specify camera body (RED V-Raptor, ARRI ALEXA), lens (anamorphic, 85mm prime), focal length, aperture (f/1.4 for shallow DOF), and aspect ratio (21:9 for cinematic).
• Video models: Describe the motion and camera movement explicitly. "Slow dolly forward through a foggy forest, golden light filtering through trees, handheld subtle shake"

Execution Flow

Step 1: Parse the Request

From $ARGUMENTS, determine:

• Asset type: image, video, edit, speech, cinema, motion-control, or app
• Subject/prompt: what to generate
• Model preference: if user specified a model name, match to the model tables above
• Input files: if user references a local file path for editing/animation

If asset type is ambiguous, default to image.

Step 2: Open Higgsfield Tab

1. Call tabs_context_mcp to check if a higgsfield.ai tab is already open
2. If found: switch to that tab
3. If not found: call tabs_create_mcp to open a new tab
4. Navigate to the target URL based on generation type + model:

URL routing:

• Image: https://higgsfield.ai/image/{model_slug} (e.g., /image/gpt-1.5, /image/seedream-4.5)
• Video (text-to-video): https://higgsfield.ai/create/video (then select model from dropdown)
• Video (image-to-video with DoP): https://higgsfield.ai/image/dop
• Edit / Inpaint: https://higgsfield.ai/edit
• Video Edit: https://higgsfield.ai/video-edit
• Speech: https://higgsfield.ai/create/speech
• Cinema Studio: https://higgsfield.ai/cinema-studio
• Motion Control: https://higgsfield.ai/create/motion-control
• Upscale: https://higgsfield.ai/library/upscale
• Mini Apps: https://higgsfield.ai/app/{slug} (e.g., /app/face-swap, /app/relight)

Step 3: Verify Page State

1. Call read_page to get the accessibility tree
2. Verify: correct page loaded, no login modal/redirect
3. If login page or auth modal detected: STOP — tell user "Please log in to higgsfield.ai in Chrome first, then retry."
4. Call take_screenshot for visual confirmation of page state

Step 4: Upload Source Files (if needed)

If the request involves editing, image-to-video, or reference images:

1. Look for an upload area/button on the page via read_page or find
2. Use upload_image to upload the source file from the local path
3. Verify the upload completed by checking the page state

Step 5: Configure Generation

1. Enter prompt: Find the prompt text area using find or read_page, then use form_input or computer (type action) to enter the crafted prompt
2. Select model (if needed): If the URL didn't route directly to the correct model:
   • Find the model selector dropdown
   • Click to open it
   • Find and click the target model name
3. Set options (optional): If user specified preferences:
   • Aspect ratio: find and click the ratio selector (1:1, 16:9, 9:16, 4:3, 21:9)
   • Quality: find and click quality setting (1K, 2K, 4K where available)
   • Duration (video): set duration if the model supports it
   • Batch count: set number of outputs if specified

Step 6: Generate

1. Find the "Generate" button (typically yellow, labeled "Generate +1" or similar)
2. Click it using computer (click action)
3. Call take_screenshot to confirm generation has started
4. Look for loading indicators, progress bars, or "Generating..." text

Step 7: Wait for Result

Poll for completion using appropriate intervals:

• Images: Check every 5-10 seconds, timeout at 120 seconds
• Video: Check every 10-15 seconds, timeout at 300 seconds
• Cinema Studio: Check every 15-20 seconds, timeout at 600 seconds

Polling method:

1. Use javascript_tool to check DOM for result elements (new images/videos appearing)
2. OR use read_page to detect result/download elements
3. OR use take_screenshot to visually check for completed results
4. Look for: new image thumbnails, video players, download buttons, or the generation spinner disappearing

If timeout reached: inform user the generation is still processing and they can check the Higgsfield library later.

Step 8: Extract Result

Once generation is complete:

1. Use javascript_tool to query DOM for result image/video URLs (look for <img src="..."> or <video src="..."> in the result area)
2. OR find a "Download" button and extract its href
3. OR use take_screenshot of the result area to show the user visually
4. Return the URL(s) and/or screenshot to the user

Step 9: Present Results

Show results and offer follow-ups based on what was generated:

After image generation:

• "Animate this into a video?" (→ DoP or other image-to-video model)
• "Upscale to 4K?" (→ Upscale tool)
• "Edit this result?" (→ Inpaint/Edit)
• "Try a different model?"

After video generation:

• "Enhance this video?" (→ Sora 2 Enhancer)
• "Try a different model?"
• "Generate another with different motion?"

After speech/avatar:

• "Try different emotion/tone?"
• "Lip sync to different audio?"

Error Handling

Examples

Basic Image

/higgsfield a cyberpunk city at night with neon signs

-> Navigates to /image/gpt-1.5, enters prompt, clicks Generate, returns result

Specific Model

/higgsfield seedream photorealistic portrait of a woman in golden hour

-> Navigates to /image/seedream-4.5, enters prompt, returns 4K photorealistic result

Fast Draft

/higgsfield z-image quick concept sketch of a robot

-> Navigates to /image/z-image, generates in 1-3 seconds

Text-to-Video

/higgsfield video of ocean waves crashing on a rocky shore at sunset

-> Navigates to /create/video, selects Sora 2, enters prompt, waits for result

Cinematic Video

/higgsfield kling 3.0 cinematic 4K video of a samurai walking through rain

-> Navigates to video page, selects Kling 3.0, generates 4K/60fps with audio

Image Animation

/higgsfield animate this image: /path/to/image.png

-> Navigates to /image/dop, uploads image, selects motion preset, generates video

Speech

/higgsfield speech saying "Welcome to the future of AI" with dramatic tone

-> Navigates to /create/speech, enters text with [dramatic] direction, generates

Cinema Studio

/higgsfield cinema studio shot of a forest with ARRI ALEXA 85mm anamorphic

-> Navigates to /cinema-studio, configures camera settings, generates cinematic shot

Editing

/higgsfield edit: add sunglasses to this portrait

-> Navigates to /edit, uploads image, enters edit prompt, generates result

Mini App

/higgsfield face swap my photo onto this character

-> Navigates to /app/face-swap, uploads source images, generates swap