Built and published langgraph-zod to close the gap between TypeScript's compile-time safety and the untyped world of LLM text: createGraphSchemas derives input/output/state boundary schemas from LangGraph channel definitions, a streaming-first SSE validator checks graph.streamEvents() payloads in flight at under 1ms per chunk, a framework-agnostic AgentBoundary wraps any async function or generator, adapters serve Next.js App Router, Hono, and Express routes, and OpenTelemetry hooks plus Pino/Winston serializers make validation failures observable in production. Zod v3 and v4 are both supported through a safe zod-ast introspection layer.

langgraph-zod — the boundary layer for language agents

TypeScript gives you compile-time safety; LLMs operate in the untyped world of unstructured text. langgraph-zod bridges that gap by deriving production-ready runtime boundary schemas directly from your LangGraph state channel definitions.

Feature highlights

Boundary Validation (createGraphSchemas): derives inputSchema and outputSchema from state definitions — omitting internal scratchpads and keys before they leak into public API output.
Streaming-First SSE Validation: incremental SSE parser + TransformStream validates graph.streamEvents() payloads on the fly, drops malformed chunks, and handles backpressure — at <1ms per-chunk latency.
Framework-agnostic AgentBoundary: wraps any async function or async generator — LangGraph, Vercel AI SDK, AutoGen TS, or plain custom agents.
Zod-native channel reducers (zodChannel) and self-healing LLM output parsing (safeParseLLMOutput) with repair retries.
Production telemetry: OpenTelemetry spans, counters (validation.error_count, channel.failed_key), and latency histograms.
CLI: export OpenAPI 3.1 specs and diff schemas for breaking changes in CI.

All of it in a 36 KB bundle, with adapters for Next.js, Hono, and Express.