Seedream

Generate AI images with BytePlus Seedream 4.5 and Seedream 5.0 Lite — from your terminal, your AI client, or Claude Code.

Choose your path

I want to…	Use
Talk to Claude in natural language (Claude Desktop, Cursor, Windsurf)	MCP Server
Generate images from the terminal	CLI
Use slash commands inside Claude Code	Claude Code Skill

All three share the same credentials and the same underlying pipeline. The Seedream image API is synchronous — a single call returns the finished image URL(s). There is no task to poll.

Models

Model	ID	Sizes	Notes
Seedream 4.5	`seedream-4-5-251128`	2K, 4K	Default. No 1K.
Seedream 5.0 Lite	`seedream-5-0-260128`	1K, 2K, 4K, adaptive	Newer, cheaper, stronger reference consistency & layout reasoning.

Aliases 4.5 and 5.0-lite are accepted anywhere a model is specified.

Prerequisites

Python 3.10+, a local clone of this repo, and keys from the BytePlus console:

Key	Where	Used for
`ARK_API_KEY`	ModelArk → API Keys	Image generation (required)
`ARK_AK`	IAM → Access Keys	TOS upload of local reference images (optional)
`ARK_SK`	IAM → Access Keys	TOS upload of local reference images (optional)
`SEEDREAM_TOS_BUCKET`	TOS console → Buckets	Default bucket for uploads (optional)
`SEEDREAM_TOS_REGION`	—	TOS region, default `ap-southeast-1` (optional)

Plain text-to-image (and references you pass as public URLs) need only ARK_API_KEY.

🖥 MCP Server

Lets Claude Desktop, Cursor, and Windsurf generate images from natural language.

Setup

git clone https://github.com/mittulmadaan/byteplus-seedream-mcp
cd byteplus-seedream-mcp
pip install packages/seedream-sdk packages/seedream-mcp -e packages/seedream-cli
seedream auth login        # stores credentials in ~/.seedream/credentials
seedream skill install     # patches Claude Desktop + Claude Code configs automatically
# Restart your AI client

Then just ask Claude: "Generate a 4K poster of iced cold brew on a sunlit table with the headline Summer Cold Brew"

Manual config

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (Mac) or the equivalent on your OS:

{
  "mcpServers": {
    "seedream": {
      "command": "/path/to/.venv/bin/python",
      "args": ["-m", "seedream_mcp"]
    }
  }
}

Set credentials via environment variables before launching your client.

Hosted / Docker

docker build -f packages/seedream-mcp/docker/Dockerfile -t seedream-mcp .

docker run -p 8000:8000 \
  -e MCP_TRANSPORT=sse \
  -e ARK_API_KEY=<key> \
  -e ARK_AK=<ak> \
  -e ARK_SK=<sk> \
  -e MCP_AUTH_TOKEN=<strong-random-token> \
  seedream-mcp

Your BytePlus credentials stay server-side. Clients connect using MCP_AUTH_TOKEN only. Point your MCP client at http://your-host:8000/mcp (streamable HTTP transport).

How to use

Use seedream_generate for everything — it returns image URLs directly:

# Text-to-image
seedream_generate(prompt="A silver espresso machine, morning light, steam rising", size="2K")
→ { model: "seedream-4-5-251128", count: 1, images: [{ url: "https://..." }] }

# Seedream 5.0 Lite, 4K
seedream_generate(prompt="Editorial close-up portrait, sculptural hat, studio lighting",
                  model="5.0-lite", size="4K")

# Editing / reference — say what to change and what to keep
seedream_generate(
  prompt="Use Image 1. Replace the background with a snowy forest at dusk. Keep the
          subject's pose, outfit, and the product label unchanged.",
  image_urls=["https://example.com/photo.png"]
)

# Consistent multi-panel set
seedream_generate(prompt="A 4-panel comic of a cat's morning, flat pastel style",
                  sequential=true, num_images=4)

For a local reference image, upload it first:

seedream_list_tos_buckets()                                  # ask which bucket
seedream_upload_image_to_tos(file_path="/path/ref.png", bucket="<name>")
→ { url: "https://<bucket>.tos-....bytepluses.com/ref.png" }
# then pass that url in image_urls

Available tools

Tool	Description
`seedream_generate`	Start here. Text-to-image, editing, reference, sequential sets. Returns URL(s).
`seedream_upload_image_to_tos`	Upload a local image; returns a public URL.
`seedream_list_tos_buckets`	List TOS buckets so Claude can ask which to use.
`seedream_list_models`	List models, sizes, and features.
`seedream_ping`	Verify credentials and connectivity.

💻 CLI

Full-featured terminal interface.

Setup

git clone https://github.com/mittulmadaan/byteplus-seedream-mcp
cd byteplus-seedream-mcp
pip install packages/seedream-sdk packages/seedream-mcp -e packages/seedream-cli
seedream auth login

Generate an image

# Text-to-image
seedream generate -p "A serene mountain landscape at sunset with golden clouds" \
  --size 2K --output out/

# Seedream 5.0 Lite, 4K
seedream generate -p "Editorial portrait, dramatic studio lighting" \
  --model 5.0-lite --size 4K --output out/

# Editing / reference — local path auto-uploads to TOS
seedream generate -p "Replace the background with a snowy forest. Keep the pose." \
  --ref ./photo.png --bucket my-bucket --output out/

# Consistent multi-panel set
seedream generate -p "A 4-panel comic of a cat's morning, flat pastel style" \
  --sequential -n 4 --output out/

All flags

seedream generate \
  --prompt        "..."       # description / edit instruction (required)
  --model         5.0-lite    # 4.5 (default) | 5.0-lite | full model ID
  --size          2K          # 1K* | 2K | 4K | adaptive* | WxH   (*5.0 Lite only)
  --ref           <url|path>  # reference image; repeatable; local paths auto-upload
  --bucket        <name>      # TOS bucket for local --ref uploads
  --seed          42          # reproducible results
  --guidance-scale 7.5        # prompt adherence 1.0–20.0
  --sequential                # generate a consistent multi-panel set
  --num-images    4           # images in a sequential set (refs + n ≤ 15)
  --optimize-prompt           # let the API optimize the prompt
  --watermark                 # add a watermark
  --output        out/        # download the image(s) into this directory

Other commands

# Auth
seedream auth login           # store credentials interactively
seedream auth check           # verify keys are configured
seedream auth logout

# TOS object storage — upload a LOCAL file, get a public URL
seedream tos buckets                             # list buckets
seedream tos upload ./photo.png --bucket <name>  # upload, prints public URL

# Info
seedream models               # list models, sizes, features
seedream ping                 # verify credentials and connectivity

# Skill management
seedream skill install        # install into Claude Desktop + Claude Code
seedream skill check          # show install status
seedream skill uninstall

🧠 Claude Code Skill

Slash commands inside Claude Code, backed by the MCP server.

Setup

git clone https://github.com/mittulmadaan/byteplus-seedream-mcp
cd byteplus-seedream-mcp
pip install packages/seedream-sdk packages/seedream-mcp -e packages/seedream-cli
seedream auth login
seedream skill install        # copies skill file + patches Claude Code settings
# Restart Claude Code

The skill lives in skills/seedream.md with detailed prompting, editing, and model-selection guidance in skills/references/.

SDK

For building custom integrations:

from seedream import SeedreamClient

client = SeedreamClient()  # reads from env vars or ~/.seedream/credentials

# Text-to-image — synchronous, returns URLs directly
result = client.generate(
    prompt="A sunrise over mountains, warm light",
    model="4.5",
    size="2K",
)
print(result.urls)

# Editing with a reference image
result = client.generate(
    prompt="Replace the background with a beach. Keep the subject unchanged.",
    image_urls=["https://example.com/photo.png"],
    model="5.0-lite",
)

# Generate and download to a directory
result = client.generate_and_download(
    "A 4-panel comic of a cat's day", "out/",
    sequential=True, num_images=4,
)
for img in result.images:
    print(img.local_path)

Install: pip install packages/seedream-sdk (zero runtime dependencies — stdlib urllib only).

Repo Structure

packages/
  seedream-sdk/          # core: HTTP client, TOS HMAC signing, config, types (zero deps)
  seedream-mcp/          # MCP server — thin wrappers over SDK
    docker/              # Dockerfile + docker-compose
  seedream-cli/          # CLI — typer + rich
    seedream_cli/commands/
      auth.py generate.py tos.py models.py skill.py

skills/
  seedream.md            # Claude Code skill definition
  references/            # prompting, editing, sequential sets, models, assets

.github/workflows/
  ci.yml                 # lint + test (Python 3.10 / 3.11 / 3.12)
  release-sdk.yml        # → PyPI on tag sdk/v*
  release-mcp.yml        # → PyPI + Docker Hub on tag mcp/v*
  release-cli.yml        # → PyPI on tag cli/v*

Credentials

Resolution order (first non-empty value wins):

Constructor args: SeedreamClient(api_key="...", ak="...", sk="...")
Environment variables: ARK_API_KEY (or BYTEPLUS_API_KEY), ARK_AK, ARK_SK
~/.seedream/credentials — written by seedream auth login
.env file in CWD (requires python-dotenv)

Never embed credentials in client config files or Docker images. Use MCP_AUTH_TOKEN to protect hosted SSE endpoints.

Release

git tag sdk/v1.0.1 && git push --tags   # publishes seedream-sdk to PyPI
git tag mcp/v1.0.1 && git push --tags   # publishes seedream-mcp to PyPI + Docker Hub
git tag cli/v1.0.1 && git push --tags   # publishes seedream-cli to PyPI

Support

Note: Seedream model IDs may change as BytePlus ships new versions. If a call returns an unknown-model error, re-check the docs above and update packages/seedream-sdk/seedream/config.py.