extensions

Official agent loop, Tool, and Environment extensions for Aex. Every package uses the public @aexhq/brain authoring API and the same brain build pipeline; official extensions have no privileged runtime path.

package	role
@aexhq/agentloop-pi	Pi-style agent loop with parallel Tool calls
@aexhq/agentloop-codex	Codex-style agent loop with sequential Tool calls
@aexhq/tools	Model-visible Tool definitions with provisioned ESM implementations
@aexhq/env-app	Application-process Environment routing callback Tools to the author's app
@aexhq/env-aws-microvm	AWS MicroVM Environment and provider runtime

All three roles have the same authoring shape:

import { agentloop } from "@aexhq/brain";

export const support = agentloop((author) => {
  author.on.message((_message, turn) => turn.reply("Hello"));
});

brain build

Agent loop handlers synchronously choose the next durable action.

Tools and Environments meet through the capability contract. A Tool declares requires and programs against typed handles (context.exec, context.fs, …); an Environment implements capability providers (provide.exec, provide.fs, …), enforces grant policy behind them, and opts in to hosting provisioned ESM artifacts with host.esm(). Brain checks requires ⊆ provides at session create and never sees inside either half:

export const bash = tool({
  description: "Run a shell command.",
  input, output,
  requires: ["exec"],
}, (author) => {
  author.run((input, context) => context.exec.run(input.command));
});

Applications place a Tool explicitly with tool().useIn(environment) and can call extension-owned methods on the same Environment object. Callback Tools stay in the application's own process: declare their schemas with appTool(...), register the functions with appTools, and bind them to @aexhq/env-app, which routes each invocation down the app's channel.