Vulca Agent Plugin

Vulca is an agent-native visual control layer for Claude Code, Gemini CLI, and Codex Desktop/CLI. It turns fuzzy creative intent into reviewable direction cards, structured prompts, semantic layers, provider-routed image work, and L1-L5 cultural evaluation.

This plugin tracks the Vulca SDK v0.23.1 package shape.

Repository links:

• SDK / CLI / MCP server: vulca-org/vulca
• Plugin package: vulca-org/vulca-plugin
• Web platform / demo: vulcaart.art
• Directory logo: assets/vulca-icon.svg

Claude Code Install

pip install "vulca[mcp]==0.23.1"
claude plugin install vulca-org/vulca-plugin

For local development, you can validate this repository directly:

claude plugin validate .
claude --plugin-dir .

The bundled MCP configuration starts vulca-mcp from your PATH. Configure provider API keys only when you explicitly want real-provider generation, editing, or evaluation. Mock/no-cost workflows work without external provider keys.

Gemini CLI Install

Gemini CLI can install this repository as an extension. The extension bundles persistent GEMINI.md context and starts the same local Vulca MCP server.

pip install "vulca[mcp]==0.23.1"
gemini extensions install vulca-org/vulca-plugin

For local development:

gemini extensions validate .
gemini extensions link .

Gemini provider-backed workflows require your own Gemini authentication, for example GEMINI_API_KEY, and remain explicit opt-in.

Codex Desktop / CLI Install

Codex can install this repository as a plugin marketplace. The plugin bundles the same skills and local vulca-mcp server configuration for Codex Desktop and Codex CLI.

pip install "vulca[mcp]==0.23.1"
codex marketplace add https://github.com/vulca-org/vulca-plugin

For local development:

codex marketplace add .

This is a Codex-compatible marketplace package. OpenAI's official public Codex plugin directory submission flow is not documented as a public form at this time.

Skills

Skill	Purpose
/decompose	Decompose an image into semantic transparent layers.
/visual-discovery	Turn fuzzy visual intent into taste profile, direction cards, and prompt artifacts.
/visual-brainstorm	Convert a selected direction into a reviewable proposal.
/visual-spec	Resolve provider, prompt, threshold, and cost decisions into design.md.
/visual-plan	Execute planned generation/evaluation iterations from design.md.
/evaluate	Evaluate an existing image against Vulca's cultural and visual rubric.
/using-vulca-skills	Meta-skill that guides when Claude should invoke the Vulca workflow skills.

Positioning

Vulca does not host a foundation image model and does not promise one-click image quality. It coordinates provider-backed workflows through auditable artifacts:

discovery.md
taste_profile.json
direction_cards.json
proposal.md
design.md
plan.md
manifest.json
evaluation.json

Redraw and inpaint tools are available as advanced MCP workflows in the SDK. They should not be marketed as polished top-level user skills until target-aware redraw evidence has been reviewed on real images.

Privacy

Vulca works with local image files and optional external image providers. When you opt into a real provider, prompts, images, and provider metadata may leave your machine depending on that provider's configuration and terms. Keep generation, editing, and VLM-backed evaluation explicit.

See PRIVACY.md for submission-ready privacy notes and SUBMISSION.md for marketplace copy and validation commands. See GEMINI.md, gemini-extension.json, and .codex-plugin/plugin.json for Gemini CLI and Codex packaging.

License

Apache-2.0