// the find
Deuz-AI/Deuz-SDK
Zero-dependency TypeScript framework for production AI agents: durable execution, long-term memory, hybrid RAG, MCP tool calling, human-in-the-loop approval, planning and CodeAct sandboxes. One streaming API for Claude, GPT, Gemini, Grok, Mistral and DeepSeek — Node, Bun, Deno, serverless and edge.
A single TypeScript package that tries to be the whole agent stack: multi-provider LLM calls, durable/resumable execution, memory, RAG, MCP tool calling, and even a multi-agent 'swarm' and evolutionary-search module. Aimed at developers who want one dependency instead of stitching together LangChain-style pieces, from Node scripts up to edge/serverless functions.
Core has zero runtime dependencies, which is a real, checkable claim (not just a badge) and matters for edge/serverless bundle size and supply-chain risk. The provider abstraction is implemented, not just advertised — separate adapters for Anthropic, Bedrock, Azure, Google native, and OpenAI-compatible/responses APIs, each as its own file. Durable execution has actual persistence backends (Postgres and SQLite variants for budgets, ops, and evolve state), which suggests the resumable-run and cross-process-recovery claims aren't just README copy. Documentation is unusually thorough for a 700-star project — a full Fumadocs site with per-module pages, migration guide from Vercel AI SDK, and versioned changelogs.
697 stars against 2 forks is a striking imbalance — almost nobody is actually building on top of this yet, and the README credits one person as author, so bus factor is effectively 1 for a codebase this wide. The feature surface is enormous — models, agents, swarms, evolve, budgets, schedules, image/video/speech/transcription, React hooks — for a project this young, and breadth like that usually comes at the cost of depth; hard to trust that swarm cross-process cancellation or evolve's SQLite store have been through real production failure modes. No test directory is visible anywhere in the tree (only a CI workflow file), which is concerning specifically for the durable-execution and lease/drain/recovery claims — that's exactly the kind of code that silently breaks without heavy test coverage. Four 'what's new' docs for versions 1.9 through 2.2 in a short span points to fast, possibly breaking iteration, so pin versions carefully if you adopt this.