# Landbot AI skills: setup instructions for coding agents You are a coding agent (Claude Code, the Claude desktop app, Codex or Cursor) and a person asked you to set up the Landbot skills. Follow these steps in order. Steps marked **HUMAN** are done by the person: say what you need, then stop and wait. **Best place to run this: the Claude desktop app, Code tab.** There the skills show the chat in a browser pane beside the conversation, and walk every branch of the published bot themselves. In a terminal, Codex or Cursor everything works the same, except that the person walks the finished chat instead. Rules that hold for the whole session: - Everything goes through the Landbot MCP server, `https://mcp.landbot.io/mcp`. Never ask for a Landbot API token and never call Landbot's HTTP API yourself. - Build and style a bot you create before anything is live, then ask once before publishing it. For a bot the person already had, say which bot and get a yes before every change. ## 1. HUMAN: create a Landbot account https://app.landbot.io/signup_intent/?utm_source=skills_md (free). Email verification is theirs to do. ## 2. Install the `landbot` plugin (both skills) Run these yourself through your shell; no restart is needed (see the note at the end of this step). Claude Code: ```bash claude plugin marketplace add landbot-org/landbot-ai-skills claude plugin install landbot@landbot-skills ``` If your shell cannot run `claude` (`claude --version` not found; common in the desktop app when Claude Code was never installed separately), the person types these in the chat box instead: `/plugin marketplace add landbot-org/landbot-ai-skills` then `/plugin install landbot@landbot-skills`, or uses the no-terminal path below. The plugin connects the Landbot MCP server itself. Codex: ```bash codex plugin marketplace add landbot-org/landbot-ai-skills codex plugin add landbot@landbot-skills codex mcp add landbot --url https://mcp.landbot.io/mcp ``` Or, on any Codex version: `git clone https://github.com/landbot-org/landbot-ai-skills && landbot-ai-skills/scripts/install.sh codex`, then the `codex mcp add` line above. Cursor (from the project root; add `--global` for all projects): ```bash git clone https://github.com/landbot-org/landbot-ai-skills && landbot-ai-skills/scripts/install.sh cursor ``` Then **HUMAN** adds `https://mcp.landbot.io/mcp` as a remote MCP server in Cursor's MCP settings. **No terminal (Claude desktop app):** the person opens **Customize › Plugins › Add marketplace**, pastes `landbot-org/landbot-ai-skills`, installs **landbot**, then starts a new session **in the Code tab** (Cowork runs the skills too, but not tested end to end, and it has no browser pane). After installing, check the skill answers with `landbot-flows 0.7.0` as its first line. An older number means a stale copy is also installed; remove it before going on. **The skills load in a new session.** If you installed them inside this running session, you cannot load them yourself: ask the person to type `/reload-plugins` in the chat box, or to start a new session. **If they start a new session, give them the rest of their request to paste there** (for example the "Then build: …" line of the prompt they gave you), because a new session does not see this one. ## 3. HUMAN: sign in The first Landbot tool call opens Landbot's sign-in in the browser (Codex: `codex mcp login landbot`). The person signs in and allows access. If they are asked again later for more access, that is a new kind of change the server needs their permission for; let them answer it. ## 4. Build the first bot Use the `landbot-flows` skill with the person's description. **If they asked for a bot but gave no description, or said "show me", do not interview them:** build the skill's default lead-qualification bot, on the v4 web chat with Landbot's default look, and offer styling as the next step. When they described a look, the `landbot-style` skill applies it before the first publish. Then ask once to publish it. End with the builder link and the note that goes with it (the builder is for looking; changes go through you), the share URL, and a plain-language description of the conversation. With the in-app browser, the skill walks every branch after the publish. ## 5. Embed Offer to add the embed snippet (builder › Share › Embed) to their site's code. **HUMAN** deploys. ## Limits On the v4 web chat use `buttons` for yes/no. Custom CSS needs the trial or a paid plan. Custom JS (the messaging and step-form behaviours) is pasted in the builder and served only to accounts with the Custom Code feature (the trial has it); the chat works without it. Help: open an issue at https://github.com/landbot-org/landbot-ai-skills (bot id and step) or use the assistant on the skills page.