OpenHarness

OpenHarness is an open-source toolkit designed for building AI agents in code, offering developers composable and stateless primitives to create agent harnesses similar to Claude Code-like products. It provides a TypeScript SDK that enables full control over agent behavior, message history, and state management, allowing users to inspect, modify, or share state between agents as plain arrays. The platform is built on top of the Vercel AI SDK, supporting integration with model providers such as OpenAI, Anthropic, Google, or any compatible provider.

Key features of OpenHarness include composable middleware for functionalities like turn tracking, retries, context compaction, and persistence, which can be mixed and matched as needed. Developers can implement subagent hierarchies, delegating tasks to specialized child agents with background execution and Promise-like combinators. The toolkit also supports automatic two-phase context management, which involves pruning old tool results and applying LLM-powered summarization to maintain relevant context for the agents.

Tool permissions can be managed through asynchronous approval callbacks, enabling gating of tool execution in various interfaces including CLI prompts, web modals, or external services. OpenHarness also offers integration with the Model Context Protocol (MCP), allowing connection to any MCP server via standard input/output, HTTP, or SSE transport. md files.

The SDK is compatible with both web and CLI environments, providing React hooks, Vue composables, and streaming support to build agents for any runtime. js or edge runtimes. OpenHarness is released under the MIT License and developed by MaxGfeller, emphasizing its open-source nature and flexibility for developers seeking to construct advanced, customizable AI agent systems.