WaveSpeedAI

Generate Images and Video Inside OpenCode with the WaveSpeed Skill

OpenCode loads skills from .claude/skills — which is exactly where wavespeed skill install writes. One command and the agent can search the model catalog, read input schemas, and run any image, video, audio, or 3D model.

By WaveSpeedAI4 min read
Generate Images and Video Inside OpenCode with the WaveSpeed Skill

OpenCode resolves agent skills from several directories, and .claude/skills is on that list — project-local and global. That makes WaveSpeed setup a genuine one-liner, because the wavespeed CLI already installs its skill exactly there:

npm install -g @wavespeed/cli
wavespeed login            # opens the key page at wavespeed.ai, validates the paste
wavespeed skill install    # writes .claude/skills/wavespeed/SKILL.md and .agents/skills/wavespeed/SKILL.md

For an npm install, the CLI requires Node.js 18 or later. In the shell OpenCode uses, check node --version, wavespeed --version, and wavespeed status before starting a generation. The skill tells the agent how to use the CLI; OpenCode’s skill and shell permissions still control whether it can load the instructions and run commands.

No plugin, no MCP server, no OpenCode-specific packaging. The skill is the same file that serves Claude Code, Cursor, and Codex.

What the agent does with it

Every model on wavespeed.ai is one wavespeed run <model-id> command. The skill teaches the find → inspect → run pattern:

# 1. FIND a model in the live catalog
wavespeed models "seedream"
wavespeed models --type image-to-video --popular

# 2. INSPECT each model's current input schema
wavespeed run bytedance/seedream-v5.0-pro -h
wavespeed run bytedance/seedance-2.5/image-to-video -h

# 3. GENERATE an image and save it locally
wavespeed run bytedance/seedream-v5.0-pro \
  -p "a cyberpunk skyline at golden hour" \
  -i aspect_ratio="16:9" -i resolution="2k" -i output_format=png \
  --json --download ./hero.png
test -s ./hero.png

That ordering matters: agents that guess parameter names can submit the wrong inputs, so the skill makes the current schema part of the loop. Continue only if the image command reports ./hero.png in its saved field and the file check succeeds.

Local files go through the @path marker — uploaded and substituted with a hosted URL in one step:

# Edit a local image
wavespeed run bytedance/seedream-v5.0-pro/edit \
  -p "replace the background with a sunlit kitchen" \
  -i images='["@./input.jpg"]' --json

# Animate the image generated above and save the video
wavespeed run bytedance/seedance-2.5/image-to-video \
  -p "subtle parallax, gentle wind" \
  -i image=@./hero.png -i resolution=720p -i duration=5 -i output_format=mp4 \
  --json --download ./hero.mp4
test -s ./hero.mp4

Use fresh output filenames for each attempt. The JSON result gives the remote URL in outputs and any downloaded file in saved; a completed generation does not by itself prove that the local download succeeded.

Cost control

Quote both steps with their intended inputs before running them:

wavespeed price bytedance/seedream-v5.0-pro \
  -i aspect_ratio="16:9" -i resolution="2k" -i output_format=png --json

wavespeed price bytedance/seedance-2.5/image-to-video \
  -i image=@./hero.png -i resolution=720p -i duration=5 -i output_format=mp4 \
  --upload --json

wavespeed balance

Many models bill from an input — audio duration or frame count, for example. A quote that omits a billing input can be too low. Here, --upload lets the video quote inspect the local image; it uses upload quota but does not submit a generation. The amount actually charged for a run is authoritative.

Sharp edges

  • Bare paths are not uploaded. -i image=./photo.jpg sends the literal string. Use ./photo.jpg.
  • Don’t invent model ​IDs — confirm via wavespeed models or wavespeed schema <id> first.
  • Use ​--json​​ on runs so the agent can read outputs[0], and check saved before relying on a local file.
  • A lost connection is not proof that a job failed. If the error includes a prediction ID, inspect it with wavespeed show <id>. Otherwise, check wavespeed history before submitting the same paid job again.
  • Project defaults and aliases live in wavespeed.json (wavespeed init writes one).

Where things live

FAQ

Why is there no OpenCode-specific package?

OpenCode's skill search path includes .claude/skills (project and global), which is exactly where `wavespeed skill install` writes. The same file serves Claude Code, Cursor, Codex, and OpenCode with zero adaptation.

Which models can the agent use?

The full live WaveSpeed catalog — image, video, audio, and 3D. The skill teaches the agent to search it with `wavespeed models`, so newly released models work without a skill update.

Project-level or global install?

Run `wavespeed skill install` inside a project to check the skill into git for your team, or copy the same directory to ~/.claude/skills to have it in every project.

Can I control spend?

Yes — `wavespeed price <model-id> -i key=value` quotes a run before executing it, and `wavespeed balance` shows account credit.

Share