Amazon Bedrock Image Generation MCP Server

Note: This is a community-maintained fork of awslabs/mcp/bedrock-image-mcp-server with additional features and improvements. Original work by Amazon Web Services under Apache 2.0 license.

MCP server for generating and editing images using Amazon Nova Canvas, Stable Diffusion 3.5 Large, and Stability AI Image Services through Amazon Bedrock.

Which model should I use?

Three Stability AI text-to-image models form a quality ladder, all in us-west-2:

In short: generate_image_ultra for a few premium, large-format or print pieces; generate_image_sd35 when you need many good assets; generate_image_core when speed and cost dominate. Stability credits only Ultra with typography, so prefer it when the image contains text.

All quotes are Stability AI's own words, from their Bedrock launch post and their API specification (api.stability.ai/v2alpha/openapi). Credits are Stability's billing unit — on Bedrock you are billed per image by AWS, so treat them as a cost ratio, not a price.

For image-to-image, transform_image_sd35 is the only option of the four; Ultra and Core are text-to-image only.

On seeds: a fixed non-zero seed reproduces the same image reliably within a short window (verified 6/6 identical), but is best-effort rather than guaranteed — repeats separated by longer intervals occasionally differ, which appears to be Bedrock serving the request from a different backend. Use seed=0 for explicitly random output.

Related MCP server: Nano-Banana MCP Server

Features

Stability AI Text-to-Image (4 tools) — recommended

Highest quality

• Generate images with generate_image_ultra (Stable Image Ultra)

• Stability AI's flagship model: best photorealism, lighting and legible text

• Same aspect ratios and prompt length as SD3.5; higher cost per image

• Text-to-image only, png or jpeg output (no webp)

Fastest and cheapest

• Generate images with generate_image_core (Stable Image Core)

• Lowest cost and latency; ideal for drafts and iterating on concepts

• Text-to-image only, png or jpeg output (no webp)

Balanced text-to-image generation

• Generate images from text prompts with generate_image_sd35

• Supports prompts up to 10,000 characters (vs 1,024 for Nova Canvas)

• 9 aspect ratio options: 16:9, 1:1, 21:9, 2:3, 3:2, 4:5, 5:4, 9:16, 9:21

• Superior prompt adherence and detail preservation

• Seed support for reproducible results (0-4,294,967,294)

Image-to-image transformation

• Transform existing images with transform_image_sd35

• Strength parameter (0.0-1.0) controls transformation intensity

• Supports file paths and base64 image inputs

• All text-to-image parameters available

Amazon Nova Canvas (2 tools)

Text-based image generation

• Create images from text prompts with generate_image

• Customizable dimensions (320-4096px), quality options, and negative prompting

• Supports multiple image generation (1-5) in single request

• Adjustable parameters like cfg_scale (1.1-10.0) and seeded generation

Color-guided image generation

• Generate images with specific color palettes using generate_image_with_colors

• Define up to 10 hex color values to influence the image style and mood

• Same customization options as text-based generation

Stability AI Upscale Services (3 tools)

Creative upscaling

• Upscale images to 4K with AI enhancement using upscale_creative

• 20-40x upscale from low-resolution inputs (64x64 to 1MP)

• Creativity parameter (0.1-0.5) controls enhancement level

• Style preset support for specific aesthetics

Conservative upscaling

• Upscale to 4K while preserving details with upscale_conservative

• Supports inputs up to 9.4 megapixels

• Minimal alterations to original image

Fast upscaling

• Quick 4x upscaling with upscale_fast

• Fast processing for quick resolution increases

• Supports inputs from 32x32 to 1MP

Stability AI Edit Services (6 tools)

Inpainting (Generative Fill)

• Fill masked regions with AI content using inpaint_image

• Grayscale mask support (white=fill, black=preserve)

• grow_mask parameter (0-20) for edge blending

Outpainting

• Extend images beyond boundaries with outpaint_image

• Directional expansion: left, right, up, down (0-2000 pixels each)

• Creativity parameter for extension style

Search and Replace

• Find and replace objects with search_and_replace

• Automatic object detection and masking

• No manual mask required

Search and Recolor

• Recolor specific objects with search_and_recolor

• Preserves structure while changing colors

• Maintains image quality

Remove Object

• Remove unwanted objects with remove_object

• Context-aware filling of removed areas

• Seamless blending with surroundings

Remove Background

• Automatic background removal with remove_background

• Returns PNG with transparency

• Handles complex subjects (hair, transparent objects)

Stability AI Control Services (4 tools)

Sketch to Image

• Convert sketches to detailed images with sketch_to_image

• control_strength parameter (0.0-1.0)

• Preserves sketch structure while adding detail

Structure Control

• Generate images from structural guides with structure_control

• Follows edge maps and structural guidance

• control_strength for adherence level

Style Guide

• Match reference image style with style_guide

• fidelity parameter (0.0-1.0) for style matching

• Accepts prompts for content description

Style Transfer

• Transfer style between images with style_transfer

• Fine-grained control: composition_fidelity, style_strength, change_strength

• Requires init_image (content) and style_image (style reference)

Mask Creation Utilities (3 tools)

Create masks programmatically for use with inpaint_image and remove_object tools. Masks are grayscale images where white pixels indicate areas to fill/remove and black pixels indicate areas to preserve.

Rectangular Mask

• Create rectangular masks with create_rectangular_mask

• Configurable position (x, y) and size (width, height)

• Optional feathering (0-50 pixels) for soft edges

• Perfect for signs, windows, rectangular objects

Ellipse Mask

• Create elliptical/circular masks with create_ellipse_mask

• Configurable center point and radii

• Optional feathering for soft edges

• Ideal for faces, balls, wheels, organic shapes

Full Mask

• Create full white masks with create_full_mask

• Covers entire image

• Useful for testing and full-image replacement

Workspace Integration

• All images saved to user-specified workspace directories with automatic folder creation

• Support for multiple output formats (PNG, JPEG, WebP)

• Unique filename generation or custom naming

AWS Authentication

• Uses AWS profiles for secure access to Amazon Bedrock services

• Supports all AWS regions where Bedrock models are available

Prerequisites

1. Install uv from Astral or the GitHub README

2. Install Python using uv python install 3.10

3. Set up AWS credentials with access to Amazon Bedrock

   • You need an AWS account with Amazon Bedrock enabled

   • Configure AWS credentials with aws configure or environment variables

   • Ensure your IAM role/user has the required permissions (see AWS IAM Permissions below)

Installation

Configure the MCP server in your MCP client configuration (e.g., for Amazon Q Developer CLI, edit ~/.aws/amazonq/mcp.json):

Pick your region deliberately. The examples below use us-west-2, which is the only region carrying the recommended text-to-image models (Ultra, Core and SD3.5) and also serves all 13 Stability edit/upscale/control tools. Nova Canvas is not in us-west-2 — use us-east-1, eu-west-1 or ap-northeast-1 for that, and note it retires 2026-09-30. See Supported AWS Regions.

{
  "mcpServers": {
    "bedrock-image-mcp-server": {
      "command": "uvx",
      "args": ["bedrock-image-mcp-server@latest"],
      "env": {
        "AWS_PROFILE": "your-aws-profile",
        "AWS_REGION": "us-west-2",
        "FASTMCP_LOG_LEVEL": "ERROR"
      },
      "disabled": false,
      "autoApprove": []
    }
  }
}

Windows Installation

For Windows users, the MCP server configuration format is slightly different:

{
  "mcpServers": {
    "bedrock-image-mcp-server": {
      "disabled": false,
      "timeout": 60,
      "type": "stdio",
      "command": "uv",
      "args": [
        "tool",
        "run",
        "--from",
        "bedrock-image-mcp-server@latest",
        "bedrock-image-mcp-server.exe"
      ],
      "env": {
        "FASTMCP_LOG_LEVEL": "ERROR",
        "AWS_PROFILE": "your-aws-profile",
        "AWS_REGION": "us-west-2"
      }
    }
  }
}

or docker after a successful docker build -t bedrock-image-mcp-server .:

# fictitious `.env` file with AWS temporary credentials
AWS_ACCESS_KEY_ID=ASIAIOSFODNN7EXAMPLE
AWS_SECRET_ACCESS_KEY=wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
AWS_SESSION_TOKEN=AQoEXAMPLEH4aoAH0gNCAPy...truncated...zrkuWJOgQs8IZZaIv2BXIa2R4Olgk

  {
    "mcpServers": {
      "bedrock-image-mcp-server": {
        "command": "docker",
        "args": [
          "run",
          "--rm",
          "--interactive",
          "--env",
          "AWS_REGION=us-east-1",
          "--env",
          "FASTMCP_LOG_LEVEL=ERROR",
          "--env-file",
          "/full/path/to/file/above/.env",
          "bedrock-image-mcp-server:latest"
        ],
        "env": {},
        "disabled": false,
        "autoApprove": []
      }
    }
  }

NOTE: Your credentials will need to be kept refreshed from your host

Installing via Smithery

To install Amazon Bedrock Image MCP Server for Claude Desktop automatically via Smithery:

npx -y @smithery/cli install bedrock-image-mcp-server --client claude

AWS Authentication

The MCP server uses the AWS profile specified in the AWS_PROFILE environment variable. If not provided, it defaults to the "default" profile in your AWS configuration file.

"env": {
  "AWS_PROFILE": "your-aws-profile",
  "AWS_REGION": "us-west-2"
}

Make sure the AWS profile has permissions to access Amazon Bedrock and the image generation models. The MCP server creates a boto3 session using the specified profile to authenticate with AWS services. Your AWS IAM credentials remain on your local machine and are strictly used for using the Amazon Bedrock model APIs.

Usage Examples

Stability AI Text-to-Image (start here)

Highest quality

# Stable Image Ultra: final assets, best text rendering
generate_image_ultra(
    prompt="A weathered brass compass on an antique nautical chart, macro photo",
    aspect_ratio="3:2",
    output_format="png"   # png or jpeg only; webp is not supported
)

Fastest draft

# Stable Image Core: quick concepts to iterate on
generate_image_core(
    prompt="Three flat vector logo concepts for a coffee shop",
    aspect_ratio="1:1"
)

Balanced default

generate_image_sd35(
    prompt="A serene mountain landscape at sunset",
    aspect_ratio="1:1"
)

Text-to-Image with Long Prompt

# SD3.5 supports up to 10,000 character prompts
generate_image_sd35(
    prompt="A detailed cyberpunk cityscape at night with neon signs, flying cars, holographic advertisements, rain-slicked streets reflecting colorful lights, towering skyscrapers with intricate architectural details, bustling crowds of people with futuristic fashion, street vendors with glowing food stalls, and a massive digital billboard displaying animated content",
    aspect_ratio="16:9",
    negative_prompt="blurry, low quality, distorted",
    seed=42
)

Image-to-Image Transformation

# Transform an existing image
transform_image_sd35(
    prompt="Transform into a watercolor painting style",
    image="/path/to/image.jpg",
    strength=0.7,
    aspect_ratio="1:1"
)

Amazon Nova Canvas

Use these when you need exact pixel dimensions, a color palette, or multiple images per request.

Text-to-Image with Explicit Dimensions

generate_image(
    prompt="A serene mountain landscape at sunset",
    width=1024,
    height=1024
)

Color-Guided Generation

# Generate with specific color palette
generate_image_with_colors(
    prompt="A modern living room interior",
    colors=["#2C3E50", "#ECF0F1", "#E74C3C"],
    width=1280,
    height=720
)

Stability AI Upscale Services

Creative Upscaling

# Upscale with AI enhancement
upscale_creative(
    image="/path/to/low_res_image.jpg",
    prompt="A professional portrait photograph",
    creativity=0.3,
    style_preset="photographic"
)

Conservative Upscaling

# Upscale preserving original details
upscale_conservative(
    image="/path/to/image.jpg",
    prompt="Product photography"
)

Fast Upscaling

# Quick 4x upscale
upscale_fast(
    image="/path/to/image.jpg"
)

Stability AI Edit Services

Inpainting

# Fill masked region
inpaint_image(
    image="/path/to/image.jpg",
    mask="/path/to/mask.png",
    prompt="A red sports car",
    grow_mask=5
)

Outpainting

# Extend image boundaries
outpaint_image(
    image="/path/to/image.jpg",
    prompt="Continue the landscape",
    left=500,
    right=500,
    creativity=0.5
)

Search and Replace

# Replace objects without manual masking
search_and_replace(
    image="/path/to/image.jpg",
    search_prompt="old wooden chair",
    prompt="modern leather armchair"
)

Search and Recolor

# Recolor specific objects
search_and_recolor(
    image="/path/to/image.jpg",
    select_prompt="the car",
    prompt="bright red color"
)

Remove Object

# Remove unwanted objects
remove_object(
    image="/path/to/image.jpg",
    mask="/path/to/object_mask.png"
)

Remove Background

# Automatic background removal
remove_background(
    image="/path/to/image.jpg"
)

Stability AI Control Services

Sketch to Image

# Convert sketch to detailed image
sketch_to_image(
    sketch="/path/to/sketch.jpg",
    prompt="A realistic portrait of a person",
    control_strength=0.7
)

Structure Control

# Generate from structural guide
structure_control(
    control_image="/path/to/edge_map.jpg",
    prompt="A modern building facade",
    control_strength=0.8
)

Style Guide

# Match reference style
style_guide(
    reference_image="/path/to/style_ref.jpg",
    prompt="A mountain landscape",
    fidelity=0.5
)

Style Transfer

# Transfer style with fine control
style_transfer(
    init_image="/path/to/content.jpg",
    style_image="/path/to/style.jpg",
    prompt="Apply artistic style",
    composition_fidelity=0.9,
    style_strength=1.0,
    change_strength=0.9
)

AWS IAM Permissions

Your AWS IAM user or role needs the following permissions to use this MCP server:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel"
            ],
            "Resource": [
                "arn:aws:bedrock:*::foundation-model/amazon.nova-canvas-v1:0",
                "arn:aws:bedrock:*::foundation-model/stability.sd3-5-large-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-creative-upscale-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-conservative-upscale-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-fast-upscale-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-inpaint-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-outpaint-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-search-replace-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-search-recolor-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-erase-object-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-remove-background-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-control-sketch-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-control-structure-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-image-style-guide-v1:0",
                "arn:aws:bedrock:*::foundation-model/us.stability.stable-style-transfer-v1:0"
            ]
        }
    ]
}

Supported AWS Regions

Region coverage differs sharply between the model families, and no single region runs all of them. Pick your AWS_REGION based on which tools you need.

Practical consequences:

• us-west-2 is the only region where Ultra, Core and SD3.5 work, and it also covers all 13 Stability edit/upscale/control tools — so it is the best single choice. Nova Canvas is not available there.

• us-east-1 covers Nova Canvas plus the 13 Stability tools, but none of the three text-to-image models.

• If you need both SD3.5 and Nova Canvas, you will need to run two server instances with different AWS_REGION values.

The Stability AI tools are invoked through US Geo cross-region inference profiles (their model IDs carry a us. prefix), so a request sent to any of the three regions may be served from another one. The underlying in-region model IDs are not enabled for direct on-demand use.

Nova Canvas is retiring

AWS moved Nova Canvas to Legacy on 2026-03-30, with end-of-life on 2026-09-30. After that date the two Nova tools will stop working. AWS also restricts Legacy models in ways that bite before then:

• New customers cannot start using a Legacy model at all

• Existing customers may lose access after 15 days of inactivity, which surfaces as ResourceNotFoundException (see Troubleshooting)

If you rely on Nova Canvas today, plan to move to generate_image_sd35 in us-west-2.

Note: verified against both the Bedrock API (GetFoundationModel lifecycle status) and the AWS model cards. Availability changes, so check your own region with:

aws bedrock list-foundation-models --region us-west-2 \
  --query "modelSummaries[?contains(modelId,'stability') || contains(modelId,'nova-canvas')].[modelId,modelLifecycle.status]"

See the AWS regional availability by model and model lifecycle pages for the authoritative lists.

Troubleshooting

Common Issues

"Model not found" or "Access denied" errors

Problem: You receive errors indicating the model is not available or you don't have access.

Solutions:

1. Verify your AWS region supports the model you're trying to use (see Supported AWS Regions). The provided model identifier is invalid almost always means the model is not in your region — most often SD3.5, which is us-west-2 only.

2. Ensure you've requested model access in the AWS Bedrock console:

   • Go to AWS Bedrock console → Model access

   • Request access for the models you want to use

   • Wait for approval (usually instant for most models)

3. Verify your IAM permissions include bedrock:InvokeModel for the specific model ARN

"This Model is marked by provider as Legacy" (Nova Canvas)

Problem: generate_image or generate_image_with_colors fails with ResourceNotFoundException: Access denied. This Model is marked by provider as Legacy and you have not been actively using the model in the last 30 days.

Cause: AWS moved Nova Canvas to Legacy on 2026-03-30, with end-of-life on 2026-09-30. Per the AWS model lifecycle policy, existing customers may lose access to a Legacy model after 15 days of inactivity, and new customers cannot use it at all. Previously granted model access does not exempt you.

Solutions:

• Prefer generate_image_sd35 in us-west-2. This is the recommended text-to-image tool and is Active, so it is the migration path rather than a workaround.

• To keep using Nova Canvas before EOL, re-request access in the Bedrock console and invoke it at least once every 15 days.

• Note that after 2026-09-30 the two Nova tools will stop working regardless.

"Response payload size exceeds limit" (Creative Upscale)

Problem: upscale_creative fails with {"detail":"Response payload size NNNNNNNN bytes exceeds limit"}.

Cause: Bedrock's InvokeModel caps the response size, and a 4K PNG upscale exceeds it. This is an API limit, not a bug in this server.

Solution: request output_format="jpeg". Creative upscale always returns roughly a 3150x3150 image, which is ~24MB as PNG (over the cap) but ~5MB as JPEG.

Note that a smaller input does not help — the output size is fixed, so a 256x256 input fails just the same with PNG. upscale_conservative and upscale_fast are unaffected and work with PNG.

"Invalid image dimensions" errors

Problem: Image generation fails with dimension validation errors.

Solutions:

• Nova Canvas: Ensure dimensions are between 320-4096 pixels and divisible by 16

• SD3.5: Use one of the supported aspect ratios (16:9, 1:1, 21:9, 2:3, 3:2, 4:5, 5:4, 9:16, 9:21)

• Upscale services: Check input image size constraints:

  • Creative/Fast: 64x64 to 1MP

  • Conservative: 64x64 to 9.4MP

"Content filtered" responses

Problem: Your generated image is blocked by content filtering.

Solutions:

1. Review your prompt for potentially sensitive content

2. Use negative prompts to exclude problematic elements

3. Adjust your prompt to be more specific and less ambiguous

4. Try different seed values

Mask validation errors (Inpainting/Remove Object)

Problem: Mask image is rejected during inpainting or object removal.

Solutions:

1. Ensure mask is grayscale (not RGB or RGBA)

2. Verify mask dimensions exactly match the input image

3. Use white (255) for areas to fill/remove, black (0) for areas to preserve

4. Save mask as PNG or JPEG format

"Image too large" warnings (Upscaling)

Problem: Warning about input image being too large for creative upscaling.

Solutions:

1. Use upscale_conservative instead for larger images (up to 9.4MP)

2. Resize your input image to under 1MP before creative upscaling

3. Use upscale_fast for quick 4x upscaling without size restrictions

AWS credentials not found

Problem: Server fails to start with AWS credential errors.

Solutions:

1. Run aws configure to set up your credentials

2. Set AWS_PROFILE environment variable to your profile name

3. Verify credentials file exists at ~/.aws/credentials

4. For temporary credentials, ensure AWS_SESSION_TOKEN is also set

Slow image generation

Problem: Image generation takes longer than expected.

Solutions:

1. This is normal - AI image generation can take 10-60 seconds depending on:

   • Model complexity (SD3.5 and upscaling are slower)

   • Image resolution

   • AWS region latency

2. Use upscale_fast instead of creative upscaling for faster results

3. Consider using a closer AWS region

4. For Nova Canvas, reduce number_of_images parameter

File path issues

Problem: Images not found or saved to unexpected locations.

Solutions:

1. Use absolute file paths for input images

2. Specify workspace_dir parameter to control output location

3. Check that output directory has write permissions

4. Verify input image files exist and are readable

Getting Help

If you encounter issues not covered here:

1. Check the AWS Bedrock documentation

2. Review the Model Context Protocol specification

3. Open an issue on the GitHub repository

4. Check AWS service health dashboard for outages

Development

Running Tests

# Install dependencies
uv sync --dev

# Run all tests
pytest

# Run with coverage
pytest --cov=awslabs --cov-report=html

# Run specific test file
pytest tests/test_server.py

Code Quality

# Format code
ruff format .

# Lint code
ruff check .

# Type check
pyright

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Contributing

Contributions are welcome! Please see CONTRIBUTING for guidelines.