I built this because I kept trying to use Claude Agent SDK, Codex, and OpenCode programmatically, and they were too "heavy" and uncustomizable for that. I wanted a TypeScript library I could import into my own app and build agents on top of. OpenHarness is built on Vercel's AI SDK 5, so it works with any model the SDK supports. The main things it gives you: loading agent instructions from AGENTS.md files, MCP server connections for tools, subagents (agents delegating to child agents), agent skills, tool approval prompts, and auto-compaction for long conversations. There's a CLI example and a Next.js chat interface in the repo if you want to try it out. Still early -- feedback on the API and what's missing would be useful.
Eventually, there will be sets of tools for various platforms, e.g., filesystem tools for Node.js and Bun, and a set of tools specifically for building agents that use certain sandbox providers.
Fair question! ADK is a great framework for building agentic systems. OpenHarness is meant to be a set of higher-level primitives that can be used to quickly run more general and capable agents, and can easily work with existing AGENTS.md or agent skills standards. Basically, if you automated a task with Claude Code, and now you want to be able to *easily* run it programmatically. Or if you want to integrate an OpenCode-like agent into your application. OpenHarness is obviously new, but it also doesn't reinvent the wheel - it's built on top of Vercel's AI SDK!
Claude Code et al. are amazing, but they are not that great to build on top of them. That's why I created OpenHarness: a code-first, composable, and configurable SDK for building powerful, capable AI agents deeply embedded in existing applications.