Support

How can we help you?

Get answers to all your Stark-related questions with
step-by-step guides from our Support center.

Using the AI Plugins (beta)

Augment your AI workflows with specialized plugins


Five persona accessibility agents, packaged as skills and bundled with the Stark MCP server. Each skill turns Stark scan results into role-appropriate fixes and can set up new scans — all scoped to the signed-in user's Stark workspace.

Skill/Invoke

For

Owns these scans

/stark-accessibility:stark-frontend Web UI, HTML/ARIA, keyboard, focus url, code, storybook, automated-test
/stark-accessibility:stark-backend lang/metadata, SSR markup, error contracts code, automated-test
/stark-accessibility:stark-mobile iOS/Android native accessibility APIs ios, android, code
/stark-accessibility:stark-design Contrast, color, target size, tokens (design language) figma
/stark-accessibility:stark-pm Posture, prioritization, compliance status, reporting url (mostly read-only)

Skills also auto-trigger when their description matches what you're doing; the slash commands above are for explicit selection (e.g. forcing the mobile skill rather than the front-end one).

The MCP server

.mcp.json connects the remote Stark MCP server over streamable HTTP

{ "mcpServers": { "stark": { "type": "http", "url": "https://mcp.getstark.ai/mcp" } } }

Authentication is OAuth, handled by the client on first connect — no token is stored in the plugin. The server scopes every call to the signed-in user, and gates the create-* tools behind the edit:projects permission - available to any non-viewer role. The skills are behavioral guidance; the actual enforcement (tenant isolation, the permission gate, write confirmation) lives in the server and the Stark app, not in skill text.

How the shared core works

Skills don't inherit from one another, so the shared rules live in shared/core-rules.md (the single source of truth) and each skill references it via ${CLAUDE_PLUGIN_ROOT}/shared/core-rules.md. Each skill also keeps a condensed copy of the critical guardrails in its own body, so the safety-relevant rules are always in context when a persona is active — not only if the reference file gets opened.

Go back to articles