wavespeed skill

WaveSpeed

WaveSpeed runs every model on its platform. There are two ways in, and they do the same things:

• MCP tools (generate_image, generate_video, …): use them when the WaveSpeed MCP server is connected. The hosted server cannot read local files: pass public URLs, or use the CLI for files on disk.
• The wavespeed CLI: use it when it is installed, for local files (@./path), batches and scripts.

For plain image and video work, use the shortcuts. For anything else (audio, 3D, upscaling, a specific endpoint, exact parameters), use the generic run.

Images and videos: the shortcuts

Task	CLI	MCP tool
Image from a prompt (optionally with reference pictures)	wavespeed generate image "<prompt>" [--image @./ref.jpg]	generate_image
Edit, restyle, combine images	wavespeed edit image "<prompt>" --image @./photo.jpg	edit_image
Video from a prompt, a first/last frame, or references	wavespeed generate video "<prompt>" [--first-frame …] [--ref-image …]	generate_video
Edit a video	wavespeed edit video "<prompt>" --video @./clip.mp4	edit_video
Continue a video	wavespeed extend video "<prompt>" --video @./clip.mp4	extend_video

They use the same models as the WaveSpeed /chat page and pick the endpoint (text, image, reference, edit) from the inputs you give. In the prompt, refer to media as @image 1, @video 1, @audio 1, in the order given ("put the product in @image 1 into the kitchen in @image 2").

• Price first for videos: CLI --estimate --json, MCP estimate_only: true. It returns the endpoint, the exact request and the price, and submits nothing.
• Models: pass -m / model only when the user names one; a named model is never switched. Without it the default is used, or another model only when the default cannot take the inputs.
• Options: --aspect-ratio, --resolution, --duration (CLI) or the same fields (MCP). Set them only when the user asks; values are snapped to what the model supports, and every change is reported.
• CLI: add --json to read the result, --download to save files.

Default models

	Default	Alternatives (-m / model)
Image	seedream-5-5 (Seedream 5.5)	seedream-5-pro, gpt-image-2.6 (GPT Image 2.6, the model behind ChatGPT Images 2.6; strong at text and photorealism), gpt-image-2.5, nano-banana-2.1, nano-banana-pro, nano-banana-2, gpt-image-2
Video	seedance-2.5 (Seedance 2.5)	seedance-2.5-turbo, kling-4.0, wan-3.0, wan-3.0-prime, minimax-h3 (cheap and fast, native stereo audio; defaults to 768p, its counterpart of 720p), older Seedance 2.0 tiers

Video resolution is 720p unless the user asks for another (MiniMax H3: 768p). Offer 480p for a cheaper draft.

Everything else: the generic run

Every model on the platform is one wavespeed run <model-id> (MCP: run_model). The model id is always explicit.

# 1. FIND a model in the live catalog (MCP: list_models)
wavespeed models "upscale"
wavespeed models --type text-to-audio --popular

# 2. INSPECT its inputs (MCP: get_model_schema)
wavespeed run <model-id> -h

# 3. RUN it, with --json so you can read the result (MCP: run_model)
wavespeed run <model-id> -p "..." -i key=value --json

run --json returns { id, model, prompt, outputs: [url, ...], saved: [path, ...], elapsed_ms, raw }. Keep id: it is the handle for wavespeed show <id> (MCP: get_prediction) if anything is interrupted. Use the URL when the user wants a link; add --download when they need the file on disk.

# Local files: @path uploads the file and passes its URL
wavespeed run <model-id> -i image=@./photo.jpg --json
wavespeed run <model-id> -i images='["@./a.jpg","@./b.jpg"]' --json

# Or upload separately when you need the URL itself (MCP: upload_file / create_upload)
URL=$(wavespeed upload ./photo.jpg --json | jq -r .url)

# Save outputs with a template
wavespeed run <model-id> -p "..." --download "./out/{index}.{ext}"

The generic run never changes the user's prompt or inputs. The single exception is explicit: an @path value uploads that file and substitutes its hosted URL. Bare paths are never uploaded.

Project config and aliases

If wavespeed.json exists (created by wavespeed init):

• defaultModel lets wavespeed run -p "…" (no model arg) work.
• Aliases are named shortcuts that bundle model + default inputs. Run wavespeed aliases to see them; wavespeed run <alias> -h shows the resolved schema. CLI -i k=v overrides alias defaults.

Auth

• CLI: wavespeed status shows whether the user is signed in. If not, ask them to run wavespeed login (opens https://wavespeed.ai/accesskey).
• MCP: if the tools ask for authorization, the user approves the connection once (in Claude Code: /mcp, pick wavespeed).
• Never ask the user to paste an API key into the chat.

Cost and credits

• wavespeed price <id> -i k=v --json (MCP: get_price) estimates a run without generating or charging. The shortcuts have --estimate / estimate_only.
• wavespeed balance (MCP: get_balance) shows the remaining credit; wavespeed usage and wavespeed billings show what was spent.
• If a run fails with Insufficient credits (or says a top-up is required), the balance is too low: do not retry. Tell the user and give them https://wavespeed.ai/top-up. You cannot pay for them.

Pitfalls

• Don't invent model IDs. Confirm with wavespeed models or wavespeed schema <id> (MCP: list_models, get_model_schema) before a generic run.
• Local files: use @./file.jpg. Bare paths are NOT uploaded and the model will reject them. The hosted MCP server cannot read local files at all.
• wavespeed delete requires --yes when run non-interactively (that includes you).
• A run that fails with a gateway or network error (502/503/504, connection reset, timeout) does NOT mean the task failed, and the submit request is never retried automatically on those errors (only an HTTP 429 rate limit, which the API returns before any task exists, is retried), so re-running can create a duplicate paid task. Do not re-run immediately. If the error contains a task id (task_id / prediction id), keep checking it with wavespeed show <id> --json (MCP: get_prediction) until its status is terminal (completed, failed, cancelled, timeout, deleted). If no id was returned, check recent predictions with wavespeed history --json (MCP: list_predictions; add the model; default window is the last 24h) and re-run once only if none matches. A client-side wait timeout or an interrupted run does not mean the task failed: tasks can run for a long time, so do not give up after a fixed time.