Install
Drop the skills into the agent workspace
The installer adds the Zebrafish router, production playbooks, verification script, and optional hooks for Claude Code, Codex, Cursor, and GitHub Copilot.
npx zebrafish-skills installAgents that make media with taste.
Brief projects, make image/video/voice, and ship only work that passes.Zebrafish is a media workspace, API, MCP, and skill system for agents. Brief projects, make image/video/voice in Studio, run workflows, then ship only approved work.
- 01Project brief
- 02Studio output
- 03Skill pass
What it is
A media app for agents. A skill system for taste.
Zebrafish is where agent-run media work happens: projects, Studio, workflows, files, reviews, approvals, model controls, API, and MCP.
Skills are reusable agent instructions that teach an agent how to use context, make the artifact, critique it, and improve it before the work leaves the system.
In the app
Studio
Generate, edit, animate, crop, and upscale image, video, and voice outputs.
Projects
Keep briefs, campaign files, references, run history, and approvals together.
Workflows
Turn repeatable media jobs into runs with steps, costs, status, and review gates.
Library
Store approved outputs so agents and teams can fetch the right asset later.
API + MCP
Let agents create drafts, run workflows, upload files, fetch assets, and debug runs.
In the skills
Context
Product purpose, audience, brand voice, visual rules, anti-references, and constraints.
Method
Shape the brief, craft the artifact, critique the result, and polish the final.
Taste checks
Reusable criteria for hierarchy, contrast, composition, artifacts, voice, and finish.
Handoff
The skill tells the agent what good means; Zebrafish stores and ships what passes.
Agent setup
Install the playbooks. Connect the account.
Zebrafish gives coding agents a deterministic setup path: workspace skills for how to operate, OAuth for the account handoff, scoped keys for MCP/API calls, and credits before paid workflow or Studio runs.
Connect
OAuth into a scoped MCP key
The agent onboarding link opens Clerk sign-in, then returns approved users to API key setup with the workflow-authoring preset selected.
https://zebrafish.dev/api/mcpFund
Buy credits when paid runs are needed
The setup path continues to prepaid credits through Stripe Checkout today, with an agent-native payment spec ready for implementation.
/settings/spendOperate
Run with approval and spend gates
Agents read the playbook, inspect projects and workflows, ask before spend or publish, then verify delivered outputs.
get_agent_playbookOur principles
01
The app runs the work.
Studio, projects, workflows, library, approvals, API, and MCP keep agent media work inspectable.
02
Skills teach taste.
Reusable instructions turn product context, design rules, and critique criteria into agent behavior.
03
Quality gates decide.
Image, video, and voice outputs are checked before they leave the app, API, or MCP client.
Skill loop
Shape, craft, critique, polish.
Skills are not prompt snippets. They are operating procedures your agent can follow before, during, and after generation.
/shapeDefine the audience, promise, content, visual lane, and anti-goals before code.
context: PRODUCT.md + DESIGN.md
output: brief, layout, anti-goals/craftGenerate the palette, mock the direction, then build against the approved contract.
context: approved shape
output: working artifact + QA evidence/critiqueScore hierarchy, density, accessibility, copy, and slop risk before shipping.
context: finished output
output: prioritized fixes + pass/fail/polishResolve the review backlog, harden responsive behavior, and preserve the approved direction.
context: critique findings
output: verified final passQuality gates
Taste is not a vibe. It is a check your agent can run.
Project context
Brief, audience, files, references, constraints, and campaign history.
Skill context
Design rules, model choices, commands, review criteria, and anti-references.
Media review
Composition, hierarchy, contrast, artifacts, voice, and finish.
Shipping gate
Only approved outputs move to the library, API, workflow, or MCP client.
API + MCP
Built for people in the app and agents in the stack.
The app keeps the work inspectable.
Briefs, files, generated media, workflow state, approvals, and run snapshots stay together instead of disappearing into prompts.
Agents can operate it directly.
Scoped keys and the hosted MCP server let agents create drafts, run workflows, upload references, review results, and fetch assets.
Questions about the app and the skills.
What does the app do?
Zebrafish is a workspace for agent-run media work: projects and campaigns, Studio generation and editing, workflow runs, asset library, approvals, model and spend controls, plus API and MCP access.
What are skills?
Skills are reusable agent instructions. They package product context, design rules, commands, and critique criteria so an agent knows how to plan, make, judge, and improve output.
How do the app and skills fit together?
Skills teach the agent what good means. The app runs the media work, keeps files and run history together, and gates what can move into the library, API, workflow, or MCP client.
How does an agent install Zebrafish?
Run npx zebrafish-skills install in the agent workspace, then use /agent-onboarding to sign in with OAuth and create a scoped MCP/API key. The flow preserves Clerk waitlist and organization approval gates.
How is this different from another image generator?
A generator returns an output. Zebrafish gives the agent a brief, design memory, skill loop, review surface, and shipping gate so the output is judged before it is used.
Can agents call it directly?
Yes. Zebrafish is designed for web app, API, and hosted MCP workflows so agents can brief projects, run workflows, upload files, review output, and fetch approved media.
Start clean