# Agent skill and plugin

The MCP server gives your agent the tools. The skill teaches it how to use them: the order of the loop, how to read evidence grades, how to turn a tactic's acceptance criteria into tests, and what it must never claim. The skill is free and MIT-licensed; audits still need an [API key](https://liftwire.io/mcp), and the free plan is enough to try it.

## What the package contains

| Path | What it is |
|---|---|
| `skills/liftwire/SKILL.md` | The skill itself |
| `skills/liftwire/prompts/` | Walkthroughs for a pricing page, a signup flow and a whole site |
| `skills/liftwire/rulebook.md` | Published tactic names and mechanisms, for when no key is connected |
| `.claude-plugin/plugin.json`, `.mcp.json` | The Claude Code plugin: the skill plus the MCP server |
| `cursor/rules.mdc` | A Cursor project rule with the loop |
| `codex/AGENTS.snippet.md` | A section to append to your project's `AGENTS.md` for Codex |
| `windsurf/mcp_config.json` | Windsurf MCP config |

## Claude Code

The package is a plugin and a one-plugin marketplace. Add it from a local copy or from the repository it is published in, then install it:

```bash
claude plugin marketplace add <path-or-github-repo>
claude plugin install liftwire@liftwire
```

The plugin connects the MCP server for you and reads your key from the `LIFTWIRE_API_KEY` environment variable, so export it before you start Claude Code. Without the plugin, connect the server with the Claude Code snippet in [Getting started](/docs/getting-started) and copy `skills/liftwire` into `~/.claude/skills/`.

## Cursor, Codex and Windsurf

These agents do not load Claude Code plugins. Connect the server with the snippets in [Getting started](/docs/getting-started), then:

- **Cursor:** copy `cursor/rules.mdc` to `.cursor/rules/liftwire.mdc` in your project.
- **Codex:** append `codex/AGENTS.snippet.md` to your project's `AGENTS.md`.
- **Windsurf:** the server config is in `windsurf/mcp_config.json`; the loop is described in the server's own instructions and prompts.

## Without a key

If the MCP server is not connected, the skill falls back to `rulebook.md`, which lists published tactics by name, grade and mechanism only. It is regenerated from the published library and may be empty while the library is in build; the skill tells the agent to say so rather than invent tactics.

## What the skill tells your agent never to claim

The skill repeats the rule on [evidence grades](/docs/evidence-grades): no numeric result for any tactic unless it is graded measured, no promises for your site, and no claim that a change worked until `verify_change` has checked the live page.