Open Harness

Open Harness is a composable, TypeScript SDK for building AI agents with full control over message history, tool execution, and middleware. Designed for developers who want stateless primitives without framework lock-in, it provides hierarchical subagent support, context compaction, and built-in MCP integration. Built on Vercel AI SDK, it supports any model provider and runtime including Node.js, edge, and browser. Its code-first approach lets you write inspectable agent logic as plain code, making debugging straightforward. Key features include stateless agents with full message control, composable middleware (turn tracking, retry, compaction, persistence), hierarchical subagents with background execution and Promise combinators, context compaction with two-phase pruning and LLM summarization, tool permission callbacks with async approval, and MCP integration via stdio, HTTP, or SSE. It also offers on-demand SKILL.md instruction loading, automatic AGENTS.md/CLAUDE.md project context injection, built-in filesystem and bash tools, React and Vue integration hooks, and streaming support for CLI and web UIs. The SDK is open-source under MIT license and works with any model provider via Vercel AI SDK, including OpenAI, Anthropic, and Google. It is best for developer experience engineers building agent-powered IDEs, backend developers creating composable agent pipelines, and AI engineers needing fine-grained control. Unlike managed platforms like LangChain or Vercel AI SDK alone, Open Harness gives you low-level control while remaining lightweight and unopinionated about your runtime or model choice.