Adding Capabilities

You don't have to write everything your agent can do.

axon install @axon/discord        # your agent becomes a Discord bot

Run that from your agent directory. It's additive — nothing you wrote changes.

Modules are installed. Prompts are not.

A module contributes tools, routes, and boot-time setup. After installing @axon/discord, your agent has a Discord client wired into its lifecycle and typed tools on axon.tools. You configure env keys and subscribe to its hooks; you never touch its source.

A module is code: it links against the runtime, has an ABI, and belongs to the agent that installed it. So it is a dependency, declared in your config:

export default defineAgent({
    modules: ["@axon/discord"],
})

Run axon prepare after installing, to regenerate type declarations. Your editor then knows exactly what was added.

A prompt package contributes tasks — and it is not installed at all.

axon . -s @cody/eslint-scout:scout

That works without declaring anything. A prompt is content, not a capability: the first time you run one it resolves into a machine-wide cache and renders on the spot. Nothing is written to your agent, nothing is added to node_modules, and no reload happens.

Which means every agent on your machine can already run every published prompt. There is no prompts: [...] array, because a prompt was never a property of one agent.

Names never collide

A published prompt is namespaced by its package, so it cannot shadow one you wrote:

src/prompts/scout.vue        →  axon . -P scout
@cody/eslint-scout:scout     →  axon . -s @cody/eslint-scout:scout

Different namespaces by construction, rather than a precedence rule you have to keep in your head.

Going further

Installing modules is covered in full — env keys, options, hooks, contributed tools — in axon install.

When you want to publish something yourself, a prompt is two commands and no build step. Building a module is the deeper track, for when a task isn't enough and you need to add a real capability.