AEX SDK – Plano de Controle Serverless para Agentes Autônomos
aexhq/aex
Orquestre agentes autônomos serverless com SDK/CLI tipado
Ver no GitHubSDK e CLI TypeScript para orquestrar sessões de agentes autônomos em escala. Declare ambiente (modelo, prompt, skills, MCP servers, arquivos), envie e receba stream tipado de eventos mais outputs capturados. Gerenciamento automático de recursos e suporte a múltiplos provedores.
Instalação
npx convez-ai aexhq/aex1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Executar agentes autônomos com controle de ambiente declarativo
- Integrar MCP servers e skills customizadas em runs isoladas
- Capturar e fazer download de outputs estruturados de execuções
Veredito

Ferramenta sólida para orquestração serverless de agentes, com API/CLI estáveis em intenção. Ainda em desenvolvimento ativo (pré-1.0), então monitor releases antes de pinnar versão em produção.
README
A tradução para o português deste README ainda está na fila.
aex
Early development. API and CLI surface are stable in intent but may shift. Pin a version and watch the GitHub releases page until we cut
1.0.
aex is the serverless control plane for autonomous agent sessions.
Declare the agent's environment - model, prompt, skills, MCP servers,
files, output dirs - submit it, and get back a typed event stream plus
captured outputs. BYOK for provider keys; aex attempts cleanup
for tracked run resources at terminal while provider and infrastructure
retention stay under their own policies.
Install
npm install @aexhq/sdk # or: pnpm add @aexhq/sdk / yarn add @aexhq/sdk
Example
import { AgentExecutor, RunModels, Skill, McpServer } from "@aexhq/sdk";
const aex = new AgentExecutor({ apiToken: process.env.AEX_API_TOKEN! });
const github = McpServer.remote({
name: "github",
url: "https://api.githubcopilot.com/mcp/",
headers: { authorization: `Bearer ${process.env.GITHUB_TOKEN!}` },
});
const runId = await aex.submitRun({
model: RunModels.CLAUDE_HAIKU_4_5,
prompt: "Summarise Q1 revenue by region.",
skills: [await Skill.fromPath("./skills/sheet-tools", { name: "sheet-tools" })],
mcpServers: [github],
secrets: { apiKey: process.env.ANTHROPIC_API_KEY! },
});
// Listen for events as the run executes...
for await (const event of aex.stream(runId)) console.log(event.type);
// ...then download the public run record — metadata, events, and
// captured outputs — as one zip once it finishes.
await aex.wait(runId);
await aex.download(runId, { to: "./run.zip" });
Same run request from the CLI: aex run --config run.json accepts
the same run-config fields ({ model, prompt, skills, mcpServers, postHook, ... });
aex events <run-id> --follow streams, and
aex wait <run-id> [--timeout 8m] blocks until the run finishes.
Composition
- Skills, MCP servers, AGENTS.md, files are first-class. Pass
per-run bundles inline with
Skill.fromPath(...)/Skill.fromFiles(...);aex.submitRunmaterializes the bytes to content-addressable, workspace-scoped asset storage before the run lands, so the same bytes are a no-op upload on subsequent runs. MCP servers can be remote or workspace-registered. - Outputs are captured, tracked resource cleanup is attempted. Managed
runs capture files created or modified by the agent;
outputs.allowedDirscan narrow that capture to specific roots. aex attempts cleanup of tracked runtime resources at terminal.
Guides
- Quickstart — install, auth, first run
- Skills — inline, local, and catalog bundles normalized to assets
- MCP servers — remote servers, headers, workspace refs
- Run config — plain credential-free run parameters
- Product capabilities and boundaries — what aex owns, inherits, and does not support
- Credentials — secrets bundle shape per provider
- Provider/runtime capabilities — supported providers and runtime routing
- Outputs — capture rules, retention, download
- Events — typed event guards, streaming
- Cleanup — runtime-resource lifecycle
- Release notes
Contribute
The SDK, CLI, contracts, conformance helpers, user-test harness, and documentation sources are open in this repo.
- Contributor flow (fork, PR, CI expectations):
CONTRIBUTING.md - Repository layout and package commands:
pnpm-workspace.yaml - Security disclosures:
SECURITY.md
FAQ
Which providers are supported?
Anthropic, DeepSeek, OpenAI, Gemini, and Mistral are accepted by the public
submission schema. Anthropic and DeepSeek are live-verified; OpenAI, Gemini,
and Mistral are accepted but live-unverified in the generated matrix. Anthropic
is the default provider; set provider for the others. Omit runtime or pass
runtime: "managed"; every provider uses the managed runtime and the same
submitRun shape / RunnerEvent stream. See the generated capability matrix.
How do provider keys work?
BYOK. Provider keys, MCP Authorization headers, and any
auxiliary secrets travel inline with each submission as a single
secrets bundle. They live in run-scoped custody for the lifetime of the
run and aex attempts cleanup/revocation for aex-held references
at terminal. We never persist tenant provider keys as workspace-level
connections; provider-side sessions and data remain under the selected
provider account's policies.
Is there a managed-key option? Managed key — planned. Until then it's BYOK only.
Can I self-host or run aex in my cloud? No — self-hosting (running aex in your own cloud) is not supported. The supported product boundary is the hosted aex control plane. See the product boundary page.
License
Apache License 2.0. See LICENSE.