Mozilla AI has published a workflow for running coding agents through Otari, its model gateway, so developers can replace a hosted frontier model without changing how they use the agent. Otari supports the Anthropic Messages API and an OpenAI-compatible endpoint; Mozilla also offers the software as an open-source, self-hosted project.
The setup lets Claude Code use one open-weight model for its main work and another for background calls and subagents. Otari can run web searches requested by the model, while an optional sandbox provides code execution for agents that lack a local shell. Managed accounts have separate US and EU endpoints.
A workspace dashboard records requests, tokens, cost, tool calls and complete agent sessions. Tags let a developer send the same task to an open model and a frontier model, then compare the runs rather than relying on general benchmark scores. The post demonstrates the process with Kimi K3 and MiniMax M3, but does not claim they match frontier systems on every task.
Mozilla argues that open models can handle routine searches, edits and refactors when people review the result. It still recommends stronger models for difficult reasoning or work where small mistakes carry a high cost. The useful change is the common routing and measurement layer, which makes that choice testable on a team’s own workload.