# Customization Source: https://docs.ideharmony.com/customization Harmony is extensible at several layers. Before the mechanics, the boundary: > [!SECURITY] > **Extension code is not sandboxed by default.** Skills, plugins, extensions, > and hooks run with the privileges of the process that loads them. Installing > one is a decision to run someone else's code on your machine with your > credentials in scope. Treat it exactly as seriously as `npm install` of an > unfamiliar package. ## The layers | Layer | What it extends | Reach | |---|---|---| | **Skill** | What an agent can do | Invoked by an agent | | **Command** | The composer | Typed by you | | **Prompt** | Reusable instruction text | Content only | | **Extension / plugin** | The application | Application privileges | | **Hook** | Lifecycle events | Fires automatically | | **MCP** | Tools Harmony publishes to other agents | Outbound | | **ACP** | External harnesses Harmony drives | Outbound | ## Direction matters Two of these point outward and are commonly confused: - **MCP** — Harmony acts as a *server*, publishing tools to hosts like Claude Code or Codex. Those agents call into your workspace. - **ACP** — Harmony acts as a *client*, driving an external Agent Client Protocol harness. Harmony calls out. See [MCP](/docs/customization/mcp). ## Scope and precedence Configuration can be defined at project scope or user scope. Project scope is discovered per repository root and is **trust-gated** — a project file does not silently gain the ability to run things because you opened the folder. > [!IMPORTANT] > A repository is untrusted input. A project-scoped configuration file is > content written by whoever wrote the repository, not by you. ## Permission ceilings An extension cannot grant itself more authority than the permission ceiling allows. Ceilings are how a skill that only needs to read files does not implicitly get the ability to run shell commands. Where a gate exists for extension permissions, it applies at load time — which is why a newly installed extension may prompt before it is usable. ## Safe mode When something loaded at startup breaks the application, safe mode starts without optional extension code so you can remove the offender. Reach for it before deleting your configuration. ## Failure isolation A failing extension should degrade its own feature rather than take the application down. If a single extension can crash the app, that is a defect worth reporting, not expected behaviour. ## What is not claimed - Extension code is **not** sandboxed. - A checksum or version pin proves *which* code you got, not that it is safe. - Trust classes describe provenance, not intent. ## Related - [Skills](/docs/customization/skills) - [MCP](/docs/customization/mcp) - [Security overview](/docs/security)