Open Harness

OpenHarness is an open source TypeScript toolkit for building AI agent harnesses, the kind of loop behind coding agents such as Claude Code. It is built on the Vercel AI SDK and works with OpenAI, Anthropic, Google or any compatible model provider. The design centers on composable primitives. Agents are stateless, so message history is a plain array that can be inspected, modified or shared between agents. Middleware for turn tracking, retries, context compaction and persistence can be mixed as needed. Context compaction works in two phases, pruning old tool results and then summarizing with a language model. Subagent hierarchies let a parent delegate to specialized children, with background execution and Promise style combinators. Other features include tool permissions through asynchronous approval callbacks that can drive a command line prompt, a web modal or an external service; MCP integration over stdio, HTTP or SSE; on demand loading of SKILL.md instructions; and automatic injection of AGENTS.md and CLAUDE.md project context. Filesystem and shell tools are supplied through provider objects, and React hooks and Vue composables support web front ends. Installation is npm install @openharness/core, with documentation and a GitHub repository linked from the site. The toolkit is free and open source, and no pricing is published.