Custom content moderation, built around your policies.

Pre-trained moderation tools (classifiers like Hive, Rekognition, or any off-the-shelf model) ship with a fixed taxonomy, so you only get the categories the vendor decided to build. If your platform needs something different, you wait on their roadmap. Musubi flips that.

You write your own policies in plain language (your terminology, examples, and edge cases) and an LLM enforces exactly those rules instead of a generic idea of what’s bad. Every decision comes back with the policy it was checked against, the model’s reasoning, a severity level, and a confidence score, so you see exactly why the decision was made. And if a call is wrong, you fix the policy and redeploy in minutes, rather than filing a ticket and waiting for someone to retrain a model.

Whichever ones work best for you. You can pick from frontier and open-source models, then benchmark them head-to-head on your own content to see which actually performs against your policies. Smaller, faster models trade a little accuracy for speed and larger ones do the reverse, so you can set that balance per surface.

No, and that’s a deliberate design choice rather than a limitation. Musubi steers frontier LLMs with your written policies instead of training a model on your content, so there’s no dataset of your stuff being used to tune anything, and nothing you label is ever pooled into a shared model that other customers benefit from. When you want better accuracy, you get it by editing the policy text yourself and seeing the change immediately, not by handing over data for retraining.

Accurate enough to trust with real volume: in production, decisions land at 90%+ F1, and as high as 98% depending on the policy and the model you choose.

That said, accuracy depends on how clearly the policy is written and which model you run it on, which is exactly why we let you benchmark models on your own content and tune each policy instead of handing you one fixed number and hoping it fits your platform.

Yes. When your team, your leadership, or a regulator asks “why was this actioned,” you have a real answer instead of a shrug at a black box — the model’s reasoning in plain language, plus any custom outputs you want to send downstream. It also makes QA much faster: you can spot a wrong call and trace it straight to the policy language that caused it.

Yes! You have complete control over your automation level. For example, you can start by sending everything to human review, then raise the threshold and let the model auto-action the high-confidence calls as your trust grows. Or you can set platform-wide rules first, then layer per-product or per-country overrides on top, so a strict market and a lighter-touch one can run different settings off the same policies. When you want to cut queue volume you turn automation up; when something needs human eyes, it still gets them. You can change this any time.

Minutes, not sprints. You draft a policy in plain language, optimize with our tools, test it against your real past decisions to see how it would have behaved, and deploy it yourself through our UI. No engineering tickets or waiting on a vendor to add a category.

If you want to be cautious, run the new version in parallel with the current one first, compare results on real content, and roll back cleanly if it’s misbehaving. The whole point is that changing what your platform enforces shouldn’t require a release cycle.

We also support connecting via MCP and CLI, so you and your AI agent can work together to make changes incredibly quickly.

Yes to both. One policy library spans every format and 150+ languages, so you’re not maintaining a separate setup per content type or region. It’s true multimodal, too: mixed content is judged together (an image and its caption, say) so context doesn’t get lost across separate passes. You write the policy once and it applies everywhere.

Yes. Run a new version against your live or historical decisions and compare how each performs on your actual content before you promote it. Teams use this to prove out a stricter harassment definition, dial in a new market’s rules, or sanity-check a change before it touches production. It’s the difference between “we think this policy is better” and “we can show it is.”