MaxGfeller/open-harness

What it solves

OpenHarness provides a framework for building general-purpose AI agents directly in code. It simplifies the process of creating agents that can interact with the real world through tools (like filesystems and shells), maintain long-term memory via sessions, and handle complex workflows using subagents and middleware.

How it works

Built on top of Vercel's AI SDK, the project uses a modular architecture:

• Agents: Stateless executors that use a specific model and a set of tools to perform tasks.
• Sessions: Wrappers around agents that manage conversation history, handle context window compaction, and manage retries.
• Middleware: Composable functions (like withTurnTracking or withRetry) that can be applied to agent runners to modify their behavior.
• Tools: Pre-built providers for filesystem access (NodeFs) and shell execution (NodeShell), as well as support for the Model Context Protocol (MCP).
• Subagents: The ability to delegate tasks to nested agents, which can be stateless, resumable, or run in the background.

Who it’s for

Developers who want to build AI-powered CLI tools, chat applications (using React or Vue), or autonomous agents that require tool use and state management.

Highlights

• Tool Integration: Built-in support for bash and filesystem operations.
• State Management: Dedicated Session and Conversation APIs for multi-turn interactions.
• Composable Logic: Middleware system for adding features like automatic retries and context compaction.
• UI Ready: Includes dedicated packages for React and Vue 3 to integrate streaming chat interfaces.
• Extensible: Supports dynamic subagent catalogs and MCP servers.