// the find
cloudwego/eino
The ultimate LLM/AI application development framework in Go.
Eino is ByteDance's Go framework for building LLM applications, positioned as a LangChain equivalent for Go developers. It covers the full stack: component abstractions (chat models, tools, retrievers), graph-based workflow composition, and pre-built agent patterns (ReAct, DeepAgent, supervisor). Aimed at Go shops that want structured AI pipelines without reaching for Python.
The graph/compose system is the real substance here — you get DAG execution, branching, parallel nodes, checkpointing, and interrupt/resume baked in, not bolted on as an afterthought. Stream handling is automatic at the orchestration layer; individual nodes don't need to know about upstream/downstream streaming concerns. The middleware system (context reduction, tool search, summarization) addresses real production problems like runaway context windows. Test coverage badge is visible and the test file density in the directory tree looks honest — adk/ alone has dozens of test files including integration tests.
Documentation lives on cloudwego.io and is primarily in Chinese first, English second — the README links go to zh/ paths, which signals where the real audience is; English docs may lag. The eino-ext component implementations (OpenAI, Ollama, Elasticsearch, etc.) live in a separate repo, meaning you're managing two go.mod dependency trees and coordinating releases across them. ByteDance backing cuts both ways: active development is great, but the framework's roadmap is shaped by internal Douyin/TikTok use cases, not the open-source community, and that can create surprising API pivots. Go 1.18 minimum means pre-generics projects can't adopt it, but also means the generic graph API (generic_graph.go) is doing non-trivial type gymnastics that will bite you when error messages get opaque.