Choose a Tool
Ark MCP has one tool per capability. Models, variations, edits and previews
are parameters, not separate tool names. Start with ark_list_models and
ark_job_capabilities to discover your actual enabled surface.
The sidebar’s Tool Reference contains one generated page per catalog tool, with exact input/output schemas, nested types, registration requirements and execution guidance. The tables below help you choose a starting point.
Images and Video
Section titled “Images and Video”| Task | Tool | What to know |
|---|---|---|
| Generate or edit images | seedream_generate_image | ModelArk key; reference edits, point/bbox edits, variations and dry_run |
| Create video | seedance_create_task | ModelArk key; select an enabled Seedance model, then poll the provider task |
| Poll video/3D tasks | gen_task_get | Batch status checks; lightweight calls can skip output persistence |
| List or cancel/delete tasks | gen_task_list, gen_task_cancel_or_delete | Owned-task lifecycle; cancellation/deletion has provider-specific semantics |
Seedream 5.5 is the built-in image default
(dola-seedream-5-5-261028). It supports text-to-image, reference and point/bbox
editing, and transparent PNG generation without an already-transparent input.
Explicit operator defaults remain unchanged. Layer decomposition and native OCR
are not supported by the MCP integration.
Regular Seedance 2.5 is the default and supports resolution="4k" on the
same model, with no separate model or feature gate. Flash uses the same create
tool at 480p/720p without Draft. Empty bindings add Flash alongside the selected
default; explicit lists remain authoritative. Discovery does not prove account
entitlement. The models enabled on the ComfyUI node pack are not automatically
available in Ark MCP.
| Task | Tool | Credential |
|---|---|---|
| Generate full-scene audio | seed_audio_generate | Seed Speech key |
| Transcribe speech | speech_to_text | Seed Speech key |
Understanding
Section titled “Understanding”| Task | Tool | Credential |
|---|---|---|
| Understand images/video | seed_understand | ModelArk key |
| Understand/transcribe/translate audio | seed_audio_understand | ModelArk key |
Seed Audio Understanding is not the Seed Speech ASR service. It uses a
server-configured ModelArk model; do not pass an audio clip to seed_understand.
Understanding and transcription are provider operations, not free setup tests.
| Task | Tool | Requirements |
|---|---|---|
| Text/image to 3D | hyper3d_create_task | ModelArk key and BYTEPLUS_MODELARK_3D_ENABLED=true |
| Image to 3D | hitem3d_create_task | Same 3D gate |
3D results use the shared gen_task_* lifecycle. The 3D feature is disabled
by default.
Avatars
Section titled “Avatars”| Task | Tool | Requirements |
|---|---|---|
| Image + audio to avatar video | omnihuman_create_task | Vision AI key and BYTEPLUS_OMNIHUMAN_ENABLED=true |
| Poll/cancel avatars | omnihuman_get_task, omnihuman_cancel_task | Vision AI credentials and gate |
OmniHuman has its own poller and cancel tool. Avatars are disabled by default.
Media Processing
Section titled “Media Processing”VOD AI MediaKit has a separate Bearer key,
BYTEPLUS_VOD_MEDIAKIT_API_KEY. Submit its operations in the background and
poll with vod_get_task, preserving the returned operation.
| Task | Tool |
|---|---|
| Enhance/upscale video | vod_enhance_video |
| Change codec, container, resolution or bitrate | vod_transcode_video |
| Burn subtitles | vod_add_subtitles |
| Erase subtitles/on-screen text | vod_remove_subtitles |
| Separate voice and background audio | vod_separate_audio |
| Retrieve MediaKit task status/results | vod_get_task |
Do not use ModelArk gen_task_get to poll a VOD task.
Uploads and Outputs
Section titled “Uploads and Outputs”| Task | Tool |
|---|---|
| Upload local/Base64 references to TOS/S3 | media_upload |
| Renew object URLs | media_presign |
| Read stored artifacts inline | seed_media_get_artifact |
| Export stored artifacts to a local path | seed_media_export_artifact |
| Recover a provider URL after persistence failed | seed_media_persist_url |
Video references need public HTTPS URLs. media_upload requires configured
object storage and accepts local paths only over stdio. Private-network URLs
are rejected. Store results promptly because provider links expire; consult
Artifacts for durable resource access.
Asset Library
Section titled “Asset Library”The private asset library uses ModelArk IAM AK/SK, not a Bearer key:
- Group management: ark_asset_group_ensure and ark_asset_group_list.
- Create/list/update assets: ark_asset_create, ark_asset_list, ark_asset_update.
- Real-person verification: ark_asset_verification_start and ark_asset_verification_result.
- Delete: ark_asset_delete, with an additional destructive-operation gate.
Use asset://<id> references in Seedance after the asset is usable.
See Asset Library for rights, verification and lifecycle details.
Background Jobs
Section titled “Background Jobs”Use these ordinary MCP tools on clients without native task augmentation:
| Tool | Purpose |
|---|---|
| ark_job_capabilities | Discover enabled targets and their schemas |
| ark_job_submit | Start a target operation and receive an Ark job ID |
| ark_job_get | Retrieve terminal original-tool results |
| ark_job_cancel | Cooperatively cancel local worker execution |
Keep job IDs distinct from provider task IDs. Honor polling hints, inspect partial variation failures, and do not retry a potentially accepted create call automatically. Start with Your First Tool Call.
Moderation Lookup
Section titled “Moderation Lookup”Use ark_moderation_get to retrieve
reasons for an existing blocked provider record, not to screen new content.
Requires the default-off BYTEPLUS_MODELARK_MODERATION_ENABLED flag + ModelArk
AK/SK and account entitlement. Account-wide trusted stdio/local-mode loopback
only; moderation:read is excluded from shared-key HTTP. Call directly in the
foreground, not via ark_job_submit or native tasks.
Ten lookups per rolling minute per runtime; no automatic retries or result persistence. Empty/not-found never means approval. Live success remains unverified; see Moderation Lookup for eligibility and bounds.