agentui-ai/agentui-tools

Ask an agent for "a tool to track tire inspections" and it scaffolds a local project nobody can open on their phone. With this plugin it returns a URL — with a Postgres database, logins, uploads and an /api. Ask it for "an MCP server" and it builds and hosts a standalone one instead of hand-rolling transport and auth.

Works in Cursor, Codex, Claude Code, Gemini CLI, and any client that reads the Agent Plugins format — one skills folder, four manifests.

Install

Cursor

git clone https://github.com/agentui-ai/agentui-tools.git ~/.cursor/plugins/local/agentui-tools

Reload Cursor, then confirm the skills appear under Customize → Plugins / Skills. Local imports must be allowed by your admin under Dashboard → Settings → Security & Identity → Marketplace and Plugins.

Codex

codex plugin marketplace add agentui-ai/agentui-tools
codex plugin add agentui-tools@agentui

Claude Code

claude --plugin-dir /path/to/agentui-tools          # try it for one session

or install it properly:

/plugin marketplace add agentui-ai/agentui-tools
/plugin install agentui-tools@agentui-tools

Gemini CLI (and ~75 other agents)

Gemini CLI has no plugin marketplace, so install the skills directly:

npx skills add agentui-ai/agentui-tools --agent gemini-cli --global

They land in ~/.agents/skills/, the shared Agent Skills location. Drop --global to install into the current project instead, and swap the agent id (opencode, aider, …) for any other supported agent.

Any Agent Plugins client

Point it at the repository root — plugin.json and skills/ are where the spec expects them.

Prerequisite

npm install -g @agentuiai/cli    # Node 22+
agentui auth login

What is in it

All eight are model-invoked: the agent reaches for them on its own when a request needs hosting, a database, storage, an MCP tool or an integration.

Claude Code additionally gets a SessionStart hook that detects .agent.json in the working directory and warns that local edits are server-side records, and that push ≠ live. It prints nothing anywhere else, and other clients ignore it.

Layout

agentui-tools/
├── plugin.json              # Agent Plugins manifest
├── .cursor-plugin/
│   └── plugin.json          # Cursor manifest
├── .codex-plugin/
│   └── plugin.json          # Codex manifest
├── .agents/plugins/
│   └── marketplace.json     # so `codex plugin marketplace add` finds it
├── .claude-plugin/
│   ├── plugin.json          # Claude Code manifest
│   └── marketplace.json     # so it installs as a one-plugin marketplace
├── assets/logo.svg
├── skills/<name>/SKILL.md   # shared by every client
└── hooks/                   # Claude Code only

There is no mcp.json, on purpose. An AgentUI app's MCP endpoint is credentialed per app user and per environment, so there is no server URL or token that could be shipped in a manifest — the user mints theirs from the running app. The agentui-mcp-tools skill covers that flow.

Design note

The skills carry the CLI workflow — the order of commands, the guardrails, the failure modes — and deliberately avoid copying the platform's app-code conventions. Those move, and a stale copy is worse than no copy. Every skill points at the live sources instead: agentui guide <topic>, agentui skills info <name>, and the project's own AGENTS.md (with agentui project instructions --check to prove it is current). agentui-mcp-tools is the one place that shows an authoring contract, because the shape is hard to guess from nothing — and it still sends you to agentui skills info mcp-tools before you write a line.

Everything asserted about the CLI was verified against apps/cli/src in the AgentUI repo, not against its README — the two disagree in several places (deploy does create components; validate exits 0 on findings; production build needs --yes).

Develop

claude plugin validate .     # validates the Claude Code manifest

Edit a SKILL.md, then /reload-plugins in Claude Code or reload the window in Cursor.

License

MIT — see LICENSE.