The whole feature list, and an honest note about what is missing.
Feature pages are usually where products start lying — a grid of nine icons, three of which describe something that exists and six of which describe a roadmap. This one only lists what the chat can do today, because a visitor who signs up expecting a capability that is not there is a refund, not a customer.
Everything below is in the product now: the model library and mid-conversation switching, the absence of a refusal layer, Thinking mode, image and document attachments, voice input, projects with standing instructions, folders, cross-device sync with export and delete, and side-by-side answers from several models at once. The last section says what OpenRogue does not have.
One account, one dropdown, the whole library: the uncensored open models that make this product what it is, and the frontier models from the GPT, Claude, Gemini and Llama families on the tiers that include them. There are no API keys to manage, no separate billing relationships, and no tab-juggling between three assistants that each know a third of your context.
The point is not the count. It is that different models are genuinely better at different things — one writes prose with rhythm, one reasons through a specification without losing the thread, one answers in under a second — and that being able to reach all of them from one conversation is worth more than any single one of them being marginally better.
Change the model from the picker and the thread continues. The new model reads everything that came before, so you can draft at speed on a small model and escalate to a large one for the passage that matters, or move a stuck reasoning problem to a different family entirely and see whether it gets further.
This is also the cheapest way to sanity-check an answer. Ask the same question of two models with different training lineages, and where they agree you can usually stop worrying. Where they disagree, you have learned something more useful than either answer on its own.
OpenRogue adds no safety middleware of its own: no classifier sitting between your prompt and the model, no injected system message telling it to decline categories of question, no blocklist of topics. The uncensored models in the library were fine-tuned without refusal behaviour, so they answer adult, dark and controversial questions the way a reference book does — directly, without a preamble about why you might not want to know.
That is a design choice about who is in charge of the conversation, not an invitation to break the law. Frontier models reached through the higher tiers still carry their own vendors' policies; nothing here can or does remove those. If a model was built to refuse something, it will still refuse it, and the library exists precisely so you can move that question to a model that was not.
Two modes sit under the composer. Instant is the default: the model answers immediately, no throat-clearing, which is what most conversation actually wants. Thinking mode asks a reasoning-capable model to work the problem step by step before it commits to an answer, and the reasoning it produced is shown in a panel you can open and read.
It is slower, and it is worth it for a specific class of work — architecture decisions, multi-constraint planning, maths, anything where the failure mode is a confident answer built on a step that was wrong. Being able to see which step was wrong is frequently more valuable than the answer itself.
Attach an image — a screenshot, a photo, a diagram, a whiteboard — and a vision-capable model will read it and work from what it sees. On a phone you can capture straight from the camera rather than saving the picture first.
Text documents attach the same way: plain text, Markdown and CSV files are read in and become part of the conversation, which is the fastest route from “here is the thing” to “now do something with it”. Longer reference material belongs in a project instead, where it stays attached to every chat in that workspace rather than to a single message.
Dictate a prompt instead of typing it. This sounds like a convenience feature and mostly is, with one real exception: spoken prompts tend to be longer and more specific than typed ones, because people stop editing themselves for brevity. More context in, better answer out. It is also the difference between using the product on a phone and merely being able to.
A project is a named workspace with its own instructions and its own reference files. Whatever you write into the instructions applies to every conversation started inside that project — a persona, a house style, a set of facts about your codebase, the rules of a world you are writing in — without pasting it at the top of each new chat.
In practice this is how people rebuild a character from another platform, keep a research thread consistent across weeks, or make a model stop reintroducing itself every session. Files added to a project become shared context for the chats inside it, so the brief and the material travel together.
Conversations organise into folders and sync to your account, so a thread started on a laptop is open on your phone without a manual step. The web app works on any modern phone or tablet browser, and a native mobile app is in beta.
Settings holds the two controls that matter for owning your own data: export everything as a JSON file, and permanently delete all cloud-stored chats. Neither is buried behind a support ticket, and neither requires you to cancel first.
Send one prompt to several models at once and read the responses next to each other as they stream in. For decisions — which framing is stronger, which explanation is actually correct, which draft has the voice you want — comparing three answers takes about as long as reading one and is considerably more informative.
No image generation. If uncensored image generation is what you came for, a platform built around it will serve you better, and the Venice AI comparison says so directly.
No character-card library and no community persona marketplace. Projects will hold a character definition perfectly well, but there is nothing here resembling the browsable catalogue of thousands of community personas that a dedicated roleplay platform offers, and the Character.AI comparison concedes that in as many words.
No deep integration into an operating system, an inbox or an office suite. OpenRogue is a chat product with a model library behind it. If what you want is an assistant that already knows what is in your calendar, one of the big vendors' assistants is the honest recommendation.
It asks a reasoning-capable model to work through the problem in explicit steps before answering, and shows you the reasoning in a panel alongside the reply. It is slower than Instant mode and noticeably better on analysis, planning, maths and code architecture — anywhere a confident wrong answer is the thing you are trying to avoid.
Images for vision-capable models, and text documents — plain text, Markdown and CSV — which are read in and become part of the conversation. Larger reference material is better added to a project, where it stays attached to every chat in that workspace instead of one message.
Yes, with a project. Write the persona, brief or style rules into the project's instructions and every chat started inside it inherits them, with whichever model you choose behind it. That is also how most people port a character over from another platform.
Yes, in your account, so it syncs across devices. We do not train models on it and we do not sell it. The provider that answers a given prompt receives that prompt, and its own policy applies there. Settings will export the whole archive as JSON or permanently delete every cloud-stored chat whenever you want.
The web app runs in any modern phone or tablet browser, including voice input and camera capture, and a native mobile app is in beta. History syncs, so a conversation started on one device continues on another.