Skip to main content
Zap’s agent framework is built on Eve — every recipe is a portable skill directory that both human creators and AI agents treat as a self-contained executable unit. Agents read the same SKILL.md and Zap.md files a human would open, collect inputs declared in frontmatter, and run the pipeline through the same CLI commands available in a terminal. There is no separate “agent mode”; the skill directory is the interface for everyone.

​
Agent Contract

Every agent working with Zap must follow six rules:
1

Read SKILL.md first

Before taking any action, load the skill’s SKILL.md. It tells you when this skill is appropriate, what the recipe does, and where the executable frontmatter lives.
2

Treat Zap.md frontmatter as executable recipe metadata

The YAML front matter in Zap.md is the source of truth for steps, inputs, budget caps, providers, and output format. Do not invent steps that are not declared there.
3

Keep prompts in prompts/*.md

All prompt text belongs in prompts/ files. Reference them from step definitions rather than inlining prompt strings in code or in the tool call.
4

Use deterministic run_zap or zap run for creator flows

When a user asks to run a known recipe, call run_zap (or zap run from the CLI). Do not improvise undeclared steps for a production creator flow.
5

Use primitive tools only for creative development or new recipe authoring

generate_image, generate_video, generate_audio, and related primitives are for exploration and new recipe development — not for substituting the deterministic creator path.
6

Keep provider spend opt-in

Explain budget quotes before incurring cost. Stop at budget caps and require explicit approval before any expensive or identity-touching generation.

​
Root Flow

The standard creator flow follows this sequence: 
creator request
  -> select zap-<slug> skill
  -> collect declared inputs
  -> validate budget
  -> run mock or live pipeline
  -> return run id and final asset URL

​
Remote Skill Registry

Zap provides a hosted registry of published skills. Agents and the CLI can query it to discover and pull recipes: 
https://zap.wzrd.tech/api/skills
https://zap.wzrd.tech/api/skills/zap
https://zap.wzrd.tech/api/skills/zap-authoring
Use zap add <name> to copy a registry skill into your local agent/skills/ directory.

​
Agent Tools

The following tools are available to the Zap agent under agent/tools/. Each is implemented with defineTool from eve/tools and registered automatically by the Eve runtime.
ToolDescription
run_zapExecute a Zap recipe deterministically from its frontmatter and return run metadata
save_zapSave an approved Zap.md recipe as a packaged Eve skill
stitch_zapStitch generated video segments into Zap.mp4 inside the Eve sandbox using ffmpeg
edit_videoSubmit a video edit or ReViz request through the Zap provider router
generate_imageSubmit an image generation or image edit request through the Zap provider router
generate_videoSubmit a video generation or extension request through the Zap provider router
generate_audioSubmit a voice, music, or SFX request through the Zap provider router
extract_last_frameExtract the last frame from a video inside the Eve sandbox using ffmpeg
upscale_frameUpscale a frame for ExtendGen conditioning
get_run_statusReturn the current status URL for a Zap run

​
Zap Pipeline Grammar

The canonical pipeline follows: 
InitialFrame -> InitialGen -> InitialGenReViz? -> ExtendGen x N -> Zap.mp4
Prefer cheap image and keyframe iteration before committing to expensive video generation. Use GMI as the primary provider where it supports the requested video capability, and fal as fallback or for image and audio capabilities.
Skill directories follow a specific file convention. See agent/skills for the full format, frontmatter field reference, and instructions on adding registry skills.