OpenHarness SDK
A TypeScript AI agent SDK for Node.js and edge runtimes, with MIT licensing, MCP tools, and model providers through Vercel AI SDK.
OpenHarness SDK is a TypeScript toolkit for developers building their own AI agents or coding assistants in the style of Claude Code. It runs in Node.js or edge runtimes and is open source under the MIT license. You control the application code; model calls use your chosen provider through Vercel AI SDK, including OpenAI, Anthropic, and Google.
Its agents keep conversation history outside the agent itself, so your application can inspect, edit, or share messages between agents. Optional middleware handles retries, persistence, and context compaction. Compaction first removes older tool results, then uses an LLM to summarize context when needed.
Agents can work with files, execute shell commands, and call custom tools. Approval callbacks let an application ask permission before a tool runs, through a terminal prompt, web dialog, or external service. MCP support connects agents to tool servers over stdio, HTTP, or SSE.
For tasks that need several agents, it supports nested delegation, background execution, and resumable subagent sessions. Dynamic subagent catalogs let an application choose the available specialists at runtime.
React hooks and Vue composables support streaming chat interfaces, while the same core can power a CLI. The SDK also loads SKILL.md instructions on demand and brings AGENTS.md or CLAUDE.md project context into agent conversations. A separate virtual filesystem provider supports sandboxed, in-memory, or SQLite-backed file access.
