Selected work
AI Model Router (AMR)
AI Model Router connects multiple AI providers to native coding workflows while keeping the lead session and selected subagents distinct.
The problem
Different coding tasks benefit from different models, but switching providers can disrupt the working environment. A developer may want one model to coordinate a task and another to handle a bounded part of it, while retaining the tools and context of the native application. The challenge is choosing where traffic should go without treating every request as interchangeable. For Claude in particular, redirecting the entire session would change the lead agent’s connection and could interfere with the native workflows the developer intended to keep using.
What was built
I built AI Model Router to bring models from different providers into Codex and Claude Code, including Claude’s desktop experience. Its Claude integration uses a selective interception proxy to redirect only marked subagent traffic. The main session keeps its native Anthropic connection, tools, dynamic workflows, Remote Control and normal Claude subscription usage. A lead agent coordinates the assignment while selected subagents use their connected provider accounts. A guided macOS setup supports installation and updates, making routing part of the existing coding environment rather than a separate place to work.
Result or current status
AMR grew out of my own daily coding workflow. I designed and implemented it to make deliberate model selection practical when balancing capability, speed and cost for an assignment. The portfolio video demonstrates that workflow and the relationship between the coordinating session and its workers. The result is a routing tool with an explicit boundary around which requests change provider. The demonstration makes that boundary visible through the relationship between the lead session and its selected workers. Model choice remains a decision for each assignment, informed by what the work needs. Keeping that choice within the native coding environment is the central product result: the developer can coordinate work across providers while continuing to use the familiar tools and session.
Play nowStep insideWalk the Pavilion. Play chess or ping-pong.