Agent-authored plugins

If Yttri cannot do something, you do not have to find a plugin or write code: ask the agent in conversation. It writes the plugin itself and offers to install it.

How it looks

You describe the task in words: “build an integration with service X so it publishes my notes there”.
The agent writes the manifest and the code, runs the validator, and stores a draft.
A proposal card appears in the chat: what the plugin is, what it does, which rights it requests, which keys it will need.
You press Install and confirm the permissions — as with any other plugin.
You open the plugin page and enter keys or tokens in its settings.
You enable the plugin.

What the agent does not do for you

Steps 4–6 are yours, and the agent cannot bypass them:

it does not install — it only proposes via a card;
it does not grant itself rights — you confirm permissions at install;
it does not enter secrets — you type keys and tokens in the settings;
it does not enable — a plugin installs disabled; you switch it on.

Until you have gone through those steps, the plugin is not installed, holds no rights, is not enabled, and knows no keys.

The same plugin as any other

An agent-authored plugin is an ordinary js_ts_v1 one: same manifest, same sandbox, same checks. It runs in a separate process, reaches out only through the gateway, and is bound by the rights you granted. Its origin buys it no privileges.

In the plugin list it is marked as agent-authored, so you can tell it apart from what you installed yourself.

If you do not want the proposal

Dismiss the card and the draft is deleted — nothing is left in the registry or on disk.