Otari by Mozilla AI: your operating system for AI governance

Mozilla AI released Otari, an Apache 2.0 open-source AI gateway that routes every request from apps and agents through a single API to enforce per-workspace model, tool and budget rules and to let teams swap model providers via a config change. Otari blocks flagged requests before they reach the provider, checks budgets on users and application keys before each request, and offers Agent Guardrails that can refuse specific tool calls and check repository changes for coding agents. Otari's usage views and session tracing record what was called, what failed and what it cost without storing prompts, responses or tool results; Mozilla AI also offers a hosted version at otari.ai.

Expert Opinion https://blog.mozilla.ai/tag/expert-opinion/ Otari by Mozilla AI: your operating system for AI governance As AI takes on more work, organisations need clear limits on models, tools and spend. Otari, Mozilla AI’s open-source gateway, brings model choice, agent governance and cost control together, so teams can adopt AI with confidence. Your models. Your tokens. Your control. Every new tool we give an AI system lets it do more for us. A browser lets it access the web. A sandbox lets it run the code it generates. Connecting it to a company's internal systems lets it change that company's data, or expose it. Each capability raises the same questions. How far can it go? How much can it spend? When it fails, how do we find out? Most teams answer these by accident: limits end up wherever the defaults put them, or nowhere. The other problem is lock-in. Most of us built our first AI features on one or two providers. Models keep getting better and cheaper, many of them open-weight, but switching means rewriting code, re-testing prompts and redoing integrations. So we stay put, and the provider sets the prices and terms. We believe no organisation should be stuck like that. Moving to a different model or tool provider should take about as much effort as changing a config value. At Mozilla AI, we built Otari around two things: governance and choice. Our own agent platform https://octonous.com/?ref=blog.mozilla.ai kept needing the same pieces: one way to reach many models and tools, budgets and limits that hold, and a record of what ran and what it cost. Otari puts them in one place. Every request from your apps and agents goes through it, so that's where you set the rules, and where you swap providers without touching the apps. Governance: decide what your AI can do A support agent drafting ticket replies needs a model and your help docs. It doesn't need the web or code execution. A research assistant needs web search above all. In Otari you set this per workspace: allowed models, web search and code execution on or off, and required guardrail checks. Flagged requests are blocked before they reach the provider. Because the rules live in the gateway, the model can't talk its way around them. Budgets go on users and application keys, and Otari checks them before each request. Start the support agent small and raise the limit once it's earned it. For coding agents, Otari’s Agent Guardrails https://github.com/mozilla-ai/otari/blob/main/docs/agent-guardrails.md?ref=blog.mozilla.ai can refuse specific tool calls and check repository changes before the agent hands work back. Knowing what happened A finished answer can hide several model calls, a fallback and a failed tool call. Otari’s usage views show what was called, what failed and what it cost. Session tracing links all of it into one timeline, without storing prompts, responses or tool results. So you can see why something got expensive, which step keeps failing, and whether the guardrail ran. Choice: switch models without rewriting your apps Your apps reach every model through one API, so switching is a config change. Fallback rules reroute requests when a provider goes down. Tools come from Otari too, so an open-weight model gets the same web search and code execution as the proprietary model it might replace. Compare them on the whole job: accuracy, retries and how much human review the output needs. Open source The gateway is Apache 2.0. Read the code, host it yourself, contribute back. Or use the hosted version at otari.ai http://otari.ai/?ref=blog.mozilla.ai . Where to start Pick one workflow you know well. Set its models, tools and budget, decide who reviews it, and watch it for a few weeks. Then decide whether it deserves more room.