convez.ai
AgenteDesenvolvimento WebAvançado
ai-vita

Open Agent SDK

ai-vita/open-agent-sdk

SDK modular TypeScript para agentes de IA com qualquer LLM

Ver no GitHub
InicialPluginScore · de 0 a 100

SDK TypeScript open-source para construir agentes de IA agnósticos a provedor, funcionando com Claude, GPT-4, Gemini e outras LLMs. Oferece controle total, arquitetura modular e suporte a sub-agentes, cache de ferramentas e sistema de skills composável — uma alternativa livre ao Claude Agent SDK.

Instalação

$npx convez-ai ai-vita/open-agent-sdk

1 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

Estrelas
7
Forks
3
Acompanhando
0
Saúde
37%
Último commit
há 8 meses
Versão
—
Issues abertas
1
Contribuidores
1

Casos de uso

  • Construir agentes de codificação e automação terminal
  • Integrar agentes com múltiplos provedores de LLM
  • Implementar sistemas com sub-agentes e multi-step reasoning
  • Executar agentes em sandboxes locais ou cloud (E2B, Vercel)

Veredito

SDK robusto e bem-arquitetado para quem precisa de agentes multi-LLM com controle total. A modularidade e suporte a sandboxes múltiplos o tornam flexível, mas exige experiência TypeScript e conhecimento de padrões de agentes de IA.

README

A tradução para o português deste README ainda está na fila.

Open Agent SDK

Open-source TypeScript SDK for building AI agents — a provider-agnostic, modular alternative to the Claude Agent SDK.

Built on the Vercel AI SDK (v6+), so it works with any LLM: Anthropic Claude, OpenAI GPT, Google Gemini, Mistral, and more.

Why Open Agent SDK? The official Claude Agent SDK is powerful but vendor-locked and opaque. Open Agent SDK gives you the same agent loop capabilities — tool use, multi-step reasoning, sub-agents, context compaction — with full control, any LLM provider, and an install-only-what-you-need package architecture.

Features

  • Provider-agnostic — swap between Claude, GPT-4, Gemini, or any Vercel AI SDK provider
  • Modular packages — use only what you need; no mandatory cloud dependencies
  • Sandbox-agnostic — run locally, on E2B, on Vercel Firecracker, or bring your own
  • Full TypeScript — strict types, Zod schemas, generic type parameters throughout
  • Agent loop built-in — step management, budget tracking, context compaction, stop conditions
  • Sub-agent support — spawn isolated agents for parallel or delegated work
  • Skills system — composable behavior modules via the Agent Skills standard
  • Tool caching — LRU cache wrapper for any tool, out of the box

Architecture

open-agent-sdk/
├── packages/
│   ├── core/               # Agent loop, types, caching, utilities
│   ├── sandbox-local/      # Local filesystem + shell sandbox
│   ├── sandbox-e2b/        # E2B cloud sandbox
│   ├── sandbox-vercel/     # Vercel Firecracker sandbox
│   ├── cli/                # `oa` — standalone CLI coding agent
│   ├── tools/              # Standard agent tools (Bash, Read, Write, Edit, Glob, Grep, …)
│   ├── tools-web/          # Web tools (WebSearch, WebFetch) via parallel-web
│   └── skills/             # Agent Skills standard (discovery, parsing, XML injection)
└── examples/
    └── coding-agent/       # Complete coding agent example

Quick Start

pnpm install
pnpm build
pnpm test

oa — Standalone CLI Agent

A terminal coding agent you can install globally and run from any directory. Session persistence, auto-compaction, streaming, skill discovery — all built in.

# Run directly from source (dev):
AI_GATEWAY_API_KEY=... pnpm --filter @open-agent-sdk/cli start

# Or build and install globally:
pnpm --filter @open-agent-sdk/cli build
npm i -g @open-agent-sdk/cli
oa                # start or resume a session
oa --new          # fresh session
oa --model openai/gpt-4o   # use a different model

# Or compile a standalone binary (requires bun):
pnpm --filter @open-agent-sdk/cli compile
./packages/cli/dist/oa     # no runtime needed

Run the example coding agent

A full-featured terminal chatbot with session persistence, auto-compaction, and streaming — in ~100 lines of code.

# Copy .env and set AI_GATEWAY_API_KEY to your Vercel AI Gateway token
cp examples/coding-agent/.env examples/coding-agent/.env.local
AI_GATEWAY_API_KEY=... pnpm --filter @open-agent-sdk/example-coding-agent chatbot
# start a fresh session:
AI_GATEWAY_API_KEY=... pnpm --filter @open-agent-sdk/example-coding-agent chatbot --new

Run a single-task agent

AI_GATEWAY_API_KEY=... pnpm --filter @open-agent-sdk/example-coding-agent start "Count lines of code in src/"

Complete Example

import { gateway } from "ai";
import { runAgent, stepCountIs } from "@open-agent-sdk/core";
import { LocalSandbox } from "@open-agent-sdk/sandbox-local";
import { createAgentTools } from "@open-agent-sdk/tools";
import { discoverSkills, skillsToXml } from "@open-agent-sdk/skills";

// Works with any Vercel AI SDK provider — swap gateway() for openai(), google(), etc.
const model = gateway("anthropic/claude-sonnet-4.6");

const sandbox = new LocalSandbox({ cwd: process.cwd() });
const { tools } = createAgentTools(sandbox);

const skills = await discoverSkills();
const system = `You are an AI agent.\n\n${skillsToXml(skills)}`;

for await (const event of runAgent({
  model,
  tools,
  system,
  messages: "Summarize this codebase.",
  stopWhen: stepCountIs(10),
  stream: true,
})) {
  if (event.type === "text-delta") process.stdout.write(event.delta);
  if (event.type === "done") console.log(`\nDone in ${event.steps} steps.`);
}

Packages

@open-agent-sdk/core

Core interfaces and the agent loop.

  • Sandbox interface — uniform API for local and cloud execution environments
  • runAgent(options) — async generator that yields AgentEvents as the agent acts
  • stepCountIs(n), budgetExceeded(fn), composeStops(...fns) — stop conditions
  • cached(tool, name, options?) — LRU cache wrapper for any tool
  • compactConversation(config, state) — token-budget context compaction
  • middleTruncate(text, maxLength) — truncate long strings for display
import { runAgent, stepCountIs } from "@open-agent-sdk/core";

for await (const event of runAgent({ model, tools, messages: "Hello!" })) {
  if (event.type === "done") console.log(event.text);
}

@open-agent-sdk/sandbox-local

Local sandbox that runs commands via child_process and accesses the real filesystem.

import { LocalSandbox } from "@open-agent-sdk/sandbox-local";

const sandbox = new LocalSandbox({ cwd: "/my/project" });
const result = await sandbox.exec("ls -la");

@open-agent-sdk/sandbox-e2b

Cloud sandbox backed by E2B. Requires @e2b/code-interpreter as a peer dependency.

import { E2BSandbox } from "@open-agent-sdk/sandbox-e2b";

const sandbox = new E2BSandbox({ apiKey: process.env.E2B_API_KEY });

@open-agent-sdk/sandbox-vercel

Cloud sandbox backed by Vercel Firecracker. Requires @vercel/sandbox as a peer dependency.

import { VercelSandbox } from "@open-agent-sdk/sandbox-vercel";

const sandbox = new VercelSandbox({ runtime: "node22" });

@open-agent-sdk/tools

Standard agent tools built on the Sandbox interface.

ToolDescription
createBashTool(sandbox)Run shell commands with timeout and output truncation
createReadTool(sandbox)Read files with line numbers, offset/limit pagination
createWriteTool(sandbox)Write files with parent directory creation
createEditTool(sandbox)Targeted string replacement in files
createGlobTool(sandbox)Glob pattern file search
createGrepTool(sandbox)Regex content search (with ripgrep when available)
createAskUserTool()Pause agent to ask the user a question
createEnterPlanModeTool(state)Enter planning mode
createExitPlanModeTool(state)Exit planning mode with a plan
createTodoWriteTool(state)Track task progress
createTaskTool(config)Spawn sub-agents for parallel/isolated work

The createAgentTools(sandbox, config?) convenience factory creates all tools in one call:

import { createAgentTools } from "@open-agent-sdk/tools";

const { tools, planModeState, todoState } = createAgentTools(sandbox, {
  tools: { Bash: { timeout: 30_000 } },
});

@open-agent-sdk/tools-web

Web tools backed by parallel-web. Install this package when you need WebSearch or WebFetch.

import { createWebSearchTool, createWebFetchTool } from "@open-agent-sdk/tools-web";

const webSearch = createWebSearchTool();
const webFetch = createWebFetchTool();

@open-agent-sdk/cli

Standalone CLI coding agent — oa. Install globally or compile to a standalone binary.

npm i -g @open-agent-sdk/cli
oa              # interactive agent in current directory
oa --new        # fresh session
oa --help       # show all options

Features: session persistence (.session.jsonl), auto-compaction, skill discovery, configurable model via --model.

@open-agent-sdk/skills

Agent Skills standard — composable behavior modules loaded from SKILL.md files.

import { discoverSkills, skillsToXml, setupAgentEnvironment } from "@open-agent-sdk/skills";

// Discover skills from .skills/ and ~/.agents/skills/
const skills = await discoverSkills();

// Inject into system prompt as XML
const systemPrompt = `You are an AI agent.\n\n${skillsToXml(skills)}`;

// Or set up a sandbox workspace with skill files
await setupAgentEnvironment(sandbox, { skills });

Open Agent SDK vs Claude Agent SDK

Open Agent SDKClaude Agent SDK
LLM providersAny (OpenAI, Anthropic, Google, Mistral, …)Anthropic only
Open sourceYes (MIT)No
Package architectureModular — install only what you needMonolithic
SandboxLocal, E2B, Vercel, or customManaged (opaque)
Agent loopTransparent, extensibleInternal, black-box
Built onVercel AI SDK (standard)Proprietary runtime

Development

pnpm build          # Build all packages
pnpm test           # Run all tests (105 tests across 8 packages)
pnpm -r typecheck   # Type-check all packages

License

MIT

Ver o original no GitHub

Bom para

Relacionados