Get started
Getting started

Install for your assistant

Add Cue as a plugin for Claude Code, Cursor, Codex, Grok, Gemini, OpenCode, or any agent that reads .agents/skills.

The catalog of procedures can be installed as a plugin, so your assistant loads it the way it loads everything else — not as a folder of markdown you copied in.

This is the light half of Cue: the skills. It does not install the ESLint preset that fails a build when a rule is broken. For that, set the project up with cue init.

Claude Code

/plugin marketplace add wiluszdamian/cue
/plugin install cue@cue

Cursor

Add wiluszdamian/cue as a plugin marketplace, then install cue. Cursor also loads the same package as an Agent Plugin from plugins/cue/plugin.json.

Codex

codex plugin marketplace add wiluszdamian/cue

Then /plugins and install cue. Codex also reads the repo marketplace at .agents/plugins/marketplace.json.

Grok

grok plugin marketplace add wiluszdamian/cue
grok plugin install cue --trust

Gemini CLI

gemini extensions install https://github.com/wiluszdamian/cue.git

OpenCode

OpenCode does not have a marketplace this package can publish to. It already reads .opencode/skills/ and .agents/skills/:

npx skills add wiluszdamian/cue -a opencode

.agents

The same command, targeting the directory several agents share:

npx skills add wiluszdamian/cue -a universal

That lands in .agents/skills/. Codex, Cursor, Gemini CLI, OpenCode and others read it from there.

What the plugin gives the assistant

/cue is the on-ramp. A model may invoke it when you ask for tests. It routes to the rest of the catalog: /compose to write a test, /resolve-locator for a selector, /bind if the repo is not wired up yet.

The plugin starts two MCP servers. cue answers read-only point lookups — explain_rule for why a rule blocked a change and what to write instead, resolve_owner for which source of guidance governs a topic, resolve_locator for a real selector from the knowledge base rather than a guessed one, and resolve_route, resolve_api, get_evidence, get_freshness and find_knowledge for what is known about a page, an endpoint, and how far to trust it. Playwright MCP handles point-in-time browser calls.

Mapping the live app still goes through playwright-cli and /survey: an accessibility tree costs more context than the question it answers, which is why exploration is the one thing deliberately kept out of MCP.

Next