Dax Reed

Dax Raad is the creator of OpenCode and Zen, tools for AI-assisted coding and access to model deployments. His work centers on bringing models into developers’ existing workflows while supporting choices beyond a single model provider. He uses the handle @thdxr; a Latent Space interview title identifies him as “Dax Reed,” a publisher spelling variant.[5][2][1]

In his 2025 interview with Baseten, Raad recalled that browser-based coding assistance repeatedly pulled him out of his editor: he had to paste code into a conversation, wait, then copy the result back. Claude Code’s ability to manipulate files directly from the terminal was the first AI coding product that clicked for him. It solved the problem he described as “the LLM can't touch my files.” A Neovim user with experience building terminal interfaces, he said the team then explored an open-source alternative that would push terminal interaction further and accommodate other models.[5]

Raad’s argument for open source was practical: supporting many models creates a long tail of behaviors that the core team cannot encounter through its own daily use. He gave Gemini as an example—community members who used it regularly could diagnose editing-tool mistakes or API bugs that his team would miss. He explicitly rejected the idea that open source automatically produces a better product, and considered Claude Code’s closed-source approach reasonable for its tightly integrated Anthropic model experience.[5]

Raad explained an agent as an LLM given a goal and access to tools: it selects a tool, receives the result as new input, and repeats until it considers the task complete or cannot solve it. That explanation puts file access and the surrounding workflow at the center of coding assistance. The OpenCode repository describes separate built-in modes for development work and read-only analysis, called build and plan.[5][6]

Raad said recurring OpenCode issues—models stopping mid-task or returning HTTP errors—taught the team that a capable model could still provide a poor coding experience when its deployment handled tool calling badly. Zen grew out of that distinction. Rather than offer as many providers as possible, the team sought a few tested deployments with service agreements and direct relationships with the people operating them. He described Zen as usable with OpenCode or other tools, combining model flexibility with a dependable default experience.[5]

His evaluation process combined personal use with a collaborator’s tests. Raad said he would use a candidate deployment as his daily driver for a day or two to judge whether it felt usable. GosuCoder supplied private evaluations involving real codebases and feature development. According to Raad, the team also ran deployments themselves to estimate a performance ceiling, then asked providers, including Baseten, whether they could approach it. He described Zen’s intended economics as operating at breakeven: pooling users’ volume to negotiate provider discounts and passing the savings back to users.[5]

Raad criticized benchmark leadership claims that select an obscure test and evaluations built around puzzles unlike everyday software work. His priority was what users could perceive directly: responsiveness, ease of use, and a coherent product experience. As he put it, “It's just the experience around it is better.” He also argued that nondeterministic results encourage elaborate beliefs about differences between tools, comparing engineers’ model opinions to astrology. These were arguments for caution about perceived superiority, rather than claims that all models or coding tools perform identically.[5]

Before this AI coding work, Raad wrote about serverless engineering and publicly discussed SST’s development. His personal website lists “How AWS Lambda Runs Your Code,” dated October 6, 2021, and “Serverless Relational Showdown,” dated October 20, 2021, on database performance with AWS Lambda. SST’s February 27, 2023 v2 announcement identifies Adam and Dax as having discussed the rewrite process on a podcast. The team release consolidated packages, introduced a new CLI, and replaced JSON configuration with TypeScript configuration—earlier context for his work on developer tooling.[2][4]