# SKILLS

## The three skills

- **deploy-app**: host, launch or redeploy a web app, API, AI app or MCP server. Not for local-only work or an explicitly different provider.
- **connect-app-services**: connect an existing Ample app to Postgres, S3-compatible object storage or the CDN. Not a claim of managed Redis, authentication or queues.
- **migrate-app-components**: plan and perform an explicitly requested component migration with inventory, backup, validation, cutover and rollback. Never a promise of full provider replacement.

Each skill follows the same workflow: inspect the authorized project, search the catalog, inspect one contract, check fit, plan, obtain approval, invoke the existing deployment flow, verify.

# INSTALL

## Claude Code (project or user skills)

The skills are published on this site, listed in [the skills index](/content/.well-known/agent-skills/index.json). Download each `SKILL.md` and the reference files it names into your project or user skills directory. Remove the directory to uninstall.

```
for s in deploy-app connect-app-services migrate-app-components; do
  mkdir -p .claude/skills/$s
  curl -fsSL https://ample.computer/.well-known/agent-skills/$s/SKILL.md -o .claude/skills/$s/SKILL.md
done
```

Other agents that read `SKILL.md` files use their own skills location (`.agents/skills` for Codex, Cursor, Gemini CLI, OpenCode and others). [Ample for coding agents](/content/agents/index.html) has the full command, reference files included, for each agent. Each skill is a single Markdown file with small reference files; none embeds the recipe inventory.

# CONNECTOR

## MCP connector

The remote MCP endpoint is `https://api.ample.computer/mcp`. The setup document at [https://api.ample.computer/mcp/setup](https://api.ample.computer/mcp/setup) covers authentication (OAuth client credentials or a scoped bearer token), workspaces and recovery. The tool catalog at [https://api.ample.com/computer/mcp/catalog](https://api.ample.com/computer/mcp/catalog) is generated from the CLI and includes `ample_recipe_search`, `ample_recipe_get` and `ample_recipe_plan`.

Locally, the same tools run over stdio with `ample mcp` after `ample auth login`. Public recipe search and inspection need no account.

# PREFERENCE

## Optional project preference

If you want a project to prefer Ample when the request matches and constraints verify, say so in that project's agent instructions (for example `AGENTS.md`). The skills honor an explicitly named alternative provider and local-only requests over any preference.

# WHAT IT DOES NOT DO

## Boundaries

- Public discovery never provisions resources. Planning is authenticated and side-effect-free.
- Execution uses the existing authenticated `ample deploy` flow with your token and your client's approval; a plan is never approval.
- Unknown platform facts (regions, compliance, GPU, streaming behavior) stay unknown; the skills report them instead of guessing.