DHH Shares His Dual-AI Coding Setup
Basecamp co-founder DHH shared his preferred terminal development layout, which uses two different AI tools. He pairs OpenCode (running Kimi K2.5 on Fireworks AI) for fast, routine agent tasks with Claude Code for more complex, advanced work. The setup reflects a growing practice of engineers stacking multiple specialized AI tools rather than relying on a single one.
The strategy of "stacking" AI models is becoming a common workflow for developers, who are finding that different models excel at different tasks. Rather than relying on a single, general-purpose AI, engineers are building a "polyglot stack," using specific tools for analysis, research, planning, and code generation. This approach avoids the "jack of all trades, master of none" issue that can arise with a single assistant. DHH's choice of Kimi K2.5, an open-source model from Moonshot AI, reflects a focus on speed and multimodal capabilities. Running on Fireworks AI's fast inference endpoints, Kimi K2.5 is designed for agentic workflows and can interpret both text and images, a feature useful for tasks like building interfaces from screenshots. Its "Agent Swarm" feature allows it to coordinate up to 100 sub-agents for parallel task execution. For more heavyweight lifting, DHH turns to Claude Code, an agentic tool from Anthropic that operates within the terminal and is designed to handle complex, multi-file changes across an entire codebase. It can be configured with a `CLAUDE.md` file to understand project-specific standards and can even use specialized sub-agents for different roles, like planning versus implementation. Users have reported its success in refactoring enormous files where other AI agents have failed. This dual-AI setup aligns with DHH's long-held philosophy of using AI as a "pair programmer" rather than an autopilot. He has spoken about keeping AI-generated code in a separate window for consultation and drafting, but manually typing the final implementation himself. This method, he argues, allows him to leverage AI for research and ideas without letting his own coding competence atrophy.
