Seedream MCP

Your creative models. One connection.

Connect your assistant, choose a model, and build your next image or video with your existing credits.

Create an API key

Choose your starting point

Image models…

Video models…

Connect your assistant

MCP endpoint

https://seedream-5.io/api/mcp

Set SEEDREAM_API_KEY in the environment that launches your AI tool. Copy the configuration below. Replace the endpoint when testing locally.

~/.codex/config.toml

[mcp_servers.seedream]
url = "https://seedream-5.io/api/mcp"
bearer_token_env_var = "SEEDREAM_API_KEY"

Streamable HTTP · Bearer

Verification & client documentation

Codex: image generation, URL editing and results verified. Cursor and Grok Build: connection and tool discovery verified. Claude Code: connection verified. Full workflows in these three clients are pending.

New models passed production image generation, URL editing and native Codex MCP tool checks. Full AI-driven workflows across clients still need validation.

Codex Cursor Claude Code Grok

Image playground

Choose a model and settings, then copy the arguments to your connected AI assistant. Connection configuration stays the same.

Example modelAspect ratioResolutionOutput format

promptRequired

Non-empty text, up to 3000 characters. Describe the image or the changes you want.

Estimated cost… credits

Supported input parameters

Parameter	Required?	Type, values & limits
`model`	Optional	`string`seedream-5-pro Select a model ID. Omitting it uses Seedream 5 Pro. Default: seedream-5-pro
`request_id`	Required	`string` 8–128 Latin letters, digits, underscores or hyphens. Use a new ID for each new image; keep it unchanged when retrying the same task.
`prompt`	Required	`string` Non-empty text, up to 3000 characters. Describe the image or the changes you want.
`aspect_ratio`	Optional	`string`1:1 · 2:3 · 3:2 · 3:4 · 4:3 · 9:16 · 16:9 · 21:9 Sets the OUTPUT image ratio. To keep a reference ratio, choose the same ratio. auto is available only for Nano Banana 2. Default: 1:1
`resolution`	Optional	`string`1K · 2K Resolution tier. Actual pixel dimensions vary with the aspect ratio. Default: 1K
`output_format`	Optional	`string`jpeg · png Use jpeg or png, in lowercase. Default: jpeg
`image_urls`	Required for image editing only	`string[]` An array of 1–10 public HTTPS image URLs with domain names, up to 4096 characters each; use the default HTTPS port or 443. IP addresses, local paths, base64 and login-required URLs are unsupported. Omit this field for text to image.

Only the parameters listed here are supported. Unsupported values fail before submission or credit reservation. Each request creates one image.

Query tools

seedream_get_model_info: Optional model selects one model; {} returns the full catalog and default Pro options.
seedream_get_balance: No arguments: {}. Returns account credits and this key’s remaining budget.
seedream_get_task: Required task_id: the UUID returned when a task is created.

Video playground

Use the same endpoint and key. Choose Standard or Pro, then one of the three modes. Copy the arguments to your connected assistant; this example does not submit a task.

Example modelResolutionAspect ratioDuration (seconds)

Generate audioSeed (optional)

promptRequired

Estimated cost… credits

Ready-to-copy arguments

Example builder · no credits spent here

Duration (seconds): 2–30

seedream_generate_video

{
  "model": "wan-3.0-standard",
  "request_id": "generate_video_001",
  "mode": "text-to-video",
  "prompt": "A ceramic teapot on a wooden table. The camera slowly moves closer in soft morning light.",
  "aspect_ratio": "adaptive",
  "resolution": "720P",
  "duration": 5,
  "audio": true
}

Use a new request_id for each new video. Keep the same ID and arguments when retrying. Changing any parameter under that ID returns a conflict.

Call seedream_generate_video once. Poll seedream_get_task with the returned task_id every 5 seconds. Completion returns a playable video URL and resource_link. The task continues if your client disconnects. Do not resubmit a task marked SUBMISSION_UNCERTAIN; contact support.

Reference media & limits

Text mode requires prompt. Keyframe mode requires at least one of first_frame_url / last_frame_url. Reference mode requires at least one reference URL array. Keyframes and reference arrays cannot be combined. Prompt is optional in the two media modes.

Public HTTPS URLs only (no local files, base64 or sign-in). Images: up to 10, JPEG/PNG/BMP/WebP, ≤20 MiB each, sides 240–8000 px, ratio ≤8. Videos: up to 5 MP4/MOV, ≤100 MiB each, sides 240–4096 px, ratio ≤8. Audio: up to 5 WAV/MP3, ≤15 MiB each. Each video/audio is 1–15 seconds; each media type totals ≤15 seconds. Reference video plus output ≤30 seconds. URLs must remain accessible until completion; actual files are validated by the generation service.

Default: Standard · text-to-video · 720P (Pro: 1080P) · adaptive · 5s · audio on. Fixed MP4 output; pixel dimensions depend on the chosen ratio.

Supported input parameters

Parameter	Required?	Type, values & limits
`request_id`	Required	`string · 8–128 · A–Z a–z 0–9 _ -`
`model`	Optional	`wan-3.0-standard \| wan-3.0-pro` Default: wan-3.0-standard
`mode`	Optional	`text-to-video \| keyframe-to-video \| reference-to-video` Default: text-to-video
`prompt`	Required	`string · 1–20000`
`aspect_ratio`	Optional	`16:9 \| 9:16 \| 1:1 \| 4:3 \| 3:4 \| adaptive` Default: adaptive
`resolution`	Optional	Default: …
`duration`	Optional	`integer · 2–30 s` Default: 5
`audio`	Optional	`boolean · true \| false` Default: true
`seed`	Optional	`integer · 0–2147483647`
`first_frame_url / last_frame_url`	Required · First / last frame	`string · HTTPS · ≤4096`
`reference_image_urls`	Optional · Reference media	`string[] · 1–10 · HTTPS · ≤4096 / URL`
`reference_video_urls`	Reference media	`string[] · 1–5 · HTTPS · ≤4096 / URL`
`reference_audio_urls`	Reference media	`string[] · 1–5 · HTTPS · ≤4096 / URL`
`reference_video_durations / reference_audio_durations`	Optional	`number[] · 1–15 s / item · total ≤15 s`

One MP4 video per request. output_format and provider/safety overrides are unsupported. Invalid tool parameters fail before reservation. URL availability and actual media limits can also cause task failure. Optional reference duration arrays must match URL order and counts; they do not bypass actual media validation.

Your API keys

Give each AI tool its own key and spending limit. Examples & parameters

Loading…