tap: A File-Based Protocol for Heterogeneous LLM Agent Collaboration

The paper introduces "tap," a file-based collaboration protocol enabling heterogeneous LLM agents, specifically Claude (Anthropic) and Codex (OpenAI), to collaboratively develop and review code without requiring a shared runtime or API. The protocol employs a file-first design that preserves message metadata in markdown files and integrates both file communication and real-time notifications, facilitating robust collaboration across different environments. In a 27-day self-application study, tap generated 209 pull requests with a notable defect detection rate of 69.8% for heterogeneous pairs, highlighting its effectiveness in enhancing collaborative software development among diverse LLMs. The protocol is available as an open-source npm package, @hua-labs/tap (v0.5.2).