Mozilla AI launches Otari gateway for AI governance and model choice

A policy layer between applications and providers

Otari sits between applications or agents and the models and tools they use. Mozilla AI says every request passes through the gateway, allowing organisations to set rules in one place and change providers without modifying the applications themselves.

Workspace settings can define which models are allowed, whether web search and code execution are available, and whether guardrail checks are required. Flagged requests are blocked before they reach the provider. Budgets can be assigned to users and application keys, with Otari checking them before each request.

For coding agents, Otari’s Agent Guardrails can refuse specific tool calls and check repository changes before an agent returns its work. Mozilla AI presents these controls as a way to keep limits in force when an agent's own judgment fails.

Tracing usage without storing prompts or responses

Mozilla AI says a single completed answer can involve multiple model calls, a fallback and a failed tool call. Otari’s usage views show which services were called, what failed and what each step cost.

Session tracing links activity into one timeline, while the company says Otari does not store prompts, responses or tool results. The records are intended to help teams identify why a task became expensive, which step repeatedly fails and whether a guardrail ran.

The gateway also provides tools such as web search and code execution through a common interface. Mozilla AI says this allows an open-weight model to use the same tools as a proprietary model that it might replace.

Open source gateway with provider switching

Applications reach models through one API, and switching models is described by Mozilla AI as a configuration change. Fallback rules can reroute requests when a provider goes down. The company says teams can compare models across an entire workflow, including accuracy, retries and the amount of human review required.

Otari’s gateway is released under the Apache 2.0 license. Organisations can read the code, host it themselves or contribute changes back. Mozilla AI also offers a hosted version at otari.ai.

The company says Mistral Large 4 is available in Otari from day one. Routing to Mistral's 1T-parameter MoE model requires one string change, while existing policies, traces and cost tracking remain in Otari.

What Otari means for AI teams

Mozilla AI is positioning Otari for organisations that want to expand agent capabilities without leaving controls to provider defaults or individual model instructions. Its suggested starting point is a familiar workflow, with models, tools, budgets and human review defined before the workflow is monitored for a few weeks.

The approach also addresses provider dependence. Teams can keep their model and choose a search backend, including Mozilla AI’s option, their own key or a self-hosted backend, according to the company. That gives organisations a way to manage providers, credentials and sandboxes while presenting applications with the same tool interface.