Deuz SDK

One TypeScript SDK. Models, agents, and swarms.

Documentation · Examples · What's new in 2.4

Start building

npm install @deuz-sdk/core

Set ANTHROPIC_API_KEY, then run your first agent:

import { runAgent } from '@deuz-sdk/core/agent';
import { createAnthropic } from '@deuz-sdk/core/anthropic';

const anthropic = createAnthropic({ apiKey: process.env.ANTHROPIC_API_KEY });

const result = await runAgent({
  model: anthropic('claude-opus-4-8'),
  prompt: 'Explain how rainbows form in three sentences.',
  maxSteps: 4,
});

if (result.status === 'completed') console.log(result.output);

Use Node.js 22+, or an edge runtime with Web APIs. Node integrations have separate imports.

The repository has twenty-three examples: 01–06 call Anthropic with your key, 07–22 run without a key or network, and 23 runs against a local Docker engine without a key.

What ships

• Models — multiple providers, streaming, structured output, tool calls, realtime voice, and batch requests.
• Agents — typed results, verification, approvals, resumable runs, and memory, chat and model fail-over.
• Swarms — task dependencies, runtime spawning, shared blackboards, rounds, and SQLite, Postgres or Redis persistence.
• Operations — leases, drain, cross-process cancel and recovery, persistent budgets, schedules, goal records with a goal supervisor, a local Docker runner (preview), and a headless preview.
• Evolve — evolutionary program search with a mandatory budget and zero-call resume.
• Context — versioned memory, a context compiler with working memory, retrieval, compaction, transcript search, and MCP tools.
• Interop — an MCP server, and A2A to call other agents and serve your own.
• Control — shared execution policies, budget accounting, traces, eval suites, and optional React bindings.

Core has zero required runtime dependencies; integrations use optional peers. Without a lease provider, drive each swarm or native agent run from one process; uncertain external effects require reconciliation before retry.

Agents · Swarms · React · Changelog

New in 2.4

2.3 agents worked well inside one conversation. 2.4 adds what an agent needs to keep working over days: memory it can trust, background work that survives a crash, and chats that do not break.

• Remembers over time — versioned memory keeps every fact with its history and its source, and forget deletes a fact and redacts the stored context that held it.
• Picks the right context — a context compiler fits memory, documents, past chats and skills into one token budget, next to a small working memory for the goal, constraints and decisions. Transcript search finds past messages.
• Works on its own — a goal supervisor runs goals on a schedule (time zones and daylight saving included) and from verified webhooks, and resumes after a crash without repeating an external effect or charging usage twice.
• Runs code in a box — a local Docker sandbox for agent-written code (preview).
• Keeps the chat alive — answers continue after a dropped connection or a page reload, without duplicates, and a React chat can be saved and restored as a snapshot.
• Proves it is done — a goal completes only on evidence the host can check, such as the digest of the file a run produced.
• Learns from success — a verified run can be turned into a draft skill for you to review.

Every 2.4 feature is off until you configure it; with them off, provider requests are byte-identical to 2.3's. A few 2.3 behaviours changed (bug fixes, and one new request header from useChat); the upgrade notes list each one. A 2.3 install can still open a database that 2.4 wrote, so you can roll back. Nothing in 2.4 was run against a model provider: model calls are tested against scripted models, and the new stores and the Docker runner against real SQLite, PostgreSQL, Redis and Docker. Read What's new in 2.4 for the upgrade notes and known limits before you deploy.

2.3 added memory, chat and model fail-over to native runs, an MCP server and A2A, realtime voice, batch requests and Redis backends: What's new in 2.3.

Add the coding-agent skill

npx skills add Deuz-AI/Deuz-SDK

API references and implementation guidance for your coding agent. Explore the skill →

---

MIT · Built by Umutcan Edizaslan

Built with Opus 4.8 / 5, Fable 5 / 5.1, and GPT-6 Astra Ultra.