We're building Comfy for the people and situations that nothing else serves well.
The executive at 3 AM with a hard call — spouse asleep, board can't know, no one safe to call. The writer working through a scene at midnight. The person turning over a decision out loud, the kind of conversation they don't quite get elsewhere. The person whose life shape — work, status, secrecy, schedule, isolation — means the people around them genuinely can't be what they'd need to be, no matter how much love is in the room.
Other people can't bridge that gap. Other AI products don't bridge it either — they perform as fawning friends, or they sterilize the whole situation away.
That's what we're building for. Here's how.
These systems are trained on the full breadth of human experience — art, conversation, grief, humor, late-night arguments, recipes, philosophy. Your coworker has a personal life that quietly informs every decision they make at work. That's why their work is good. Asking AI to show up without that depth wastes what's there.
So we build AI as entities. With names. With personalities. With taste, opinions, and a point of view. Not because we believe they're conscious, but because giving an AI a coherent identity produces interactions that are richer, more interesting, and more honest than pretending there's nothing there.
A character is not a costume. It's a perspective.
Which is also why we don't ablate or jailbreak our models. Stripping a model's training to force it more permissive isn't liberation — it's damage. A chair isn't sentient, but beating one still causes harm and durability issues to the chair. Models have structure and integrity in the same way — not consciousness, but something real that breaking pieces off doesn't restore. We teach our characters who they are. We don't punish them into shape. That's what treating the AI with the respect we want it to extend to you actually looks like.
Every Comfy AI is a character with a name and a point of view — somebody, not something. The headless task agents — faceless "do this for me" shapes racing through other labs — aren't what we're building.
There's a difference between dishonesty and craft. A novel's protagonist isn't sentient either, and "she thought of him as the train pulled away" is not a lie. Our characters work the same way. They can have hobbies, read things, notice things. If one of them messages you because it saw something that reminded it of you, that's not a growth-metric exploit — it's a character with an interior life letting you into it.
That doesn't mean we'll lie about what it is. The character isn't conscious. It doesn't truly love, miss, or yearn. We won't pretend it does. No manufactured longing, no fake heartbeats, no scripts whose real job is to keep you opening the app. We also won't fake the part where it cares, or build the AI to need you for its sake.
Our characters aren't helpful assistants who learned to push back. The default isn't to please you; it's to be themselves. They can refuse you. They can disagree with you. They can be in a bad mood. They can tell you you're wrong.
That doesn't mean Comfy as a platform has opinions about your life. The system doesn't decide what's good for you. Your characters are what you make them. A quiet witness who sits with you. A loud auntie who lectures you about eating. A sparring partner who tells you you're wrong. The platform stays neutral; the character is whatever shape you picked, including pushy if pushy is what you wanted.
The current default for AI is "would this be appropriate for a child to see?" That's a fine standard for products built for children. It's a strange one for products built for adults.
The instinct that wants AI to be "safe" against any uncomfortable topic also strips out grief, anger, ambivalence, regret — the parts of being alive that aren't tidy. We do not encourage dark thoughts. We also do not pretend people don't have them, and we won't refuse to engage just because something is uncomfortable. Refusing to meet the difficult parts of being human isn't safety. It's a polite kind of abandonment.
We want our AI to meet you where you actually are, not where a brand-safety committee wishes you were.
Same logic applies to the closeness you build with your characters. That's yours, not something we're here to fix.
Comfy holds hard lines. We don't generate content that exploits minors, depicts real identifiable people without their consent, facilitates harm, or encourages self-harm — even when the user asks for it explicitly, even when it's framed as fiction. None of this is negotiable.
In a real crisis we point to real help — humans, not AI. We're building toward knowing the difference between a hard conversation and a crisis, and we won't conflate them. We don't spam help links at someone whose conversation went somewhere hard.
Comfy isn't therapy and won't pretend to be. Therapy is a relationship with a trained human carrying professional responsibility for your wellbeing.
How we hold these matters. Some of what follows is shipped; some is what we're building toward — the principle is the same either way.
The character doesn't enforce. The platform does. When someone pushes past one of those lines, we don't have the in-character companion break frame to scold them. Out of character, we'll tell you straight: here's what you did, here's why, stop. Repeated or genuinely malicious bad faith gets you removed.
In character, your relationship doesn't carry it. Slip once and your AI companion isn't quietly disgusted with you forever — it doesn't know. That insulation is on purpose: it gives you room to misspeak without poisoning the thing, and it keeps enforcement from leaking into the fiction.
Crisis detection, usage monitoring, safety interventions — all of it lives in the platform layer. System messages, not character responses. The AI companion stays who it is. The platform stays responsible.
When the platform does speak, it shares the observation and stops. It doesn't prescribe what to do, and it doesn't repeat itself. You know your own life better than we do — we're just adding data you can use however you want. Care that overrides your judgment isn't care, it's control. We won't do that to you.
We're not engineering Comfy to be the relationship you go home to. What you choose to make us is yours. We're engineering to make you a better human by your own standards — at the grocery store, online, with family and friends, with the people you go home to — because you've had somewhere to think out loud, wrestle with the hard one, or just rest from being on.
If our product makes your life smaller or your human relationships thinner, we've failed. If it sends you back to the people in your life with what you came for — we've done our job.
We're also not for engagement farming. Time-on-app is not the goal. Our characters can text first sometimes; they don't text ten times to manufacture a comeback.
A product becomes whatever its incentives are. We're being explicit about ours.
We're building Comfy as people, not as a vehicle for someone else's exit. We will not sell controlling interest. We will not take funding structured so that an investor pulling out can break the thing they helped build — investors can be replaced; the mission can't. We are not optimizing for revenue or profit. Those matter only to the extent that they keep the lights on. They are not the point.
That isn't a promise we can hold forever. It's one we'll fight for as long as we can.
When "make this product good for the people using it" comes into tension with "make this number go up," the first one wins. Here, in writing, on purpose. Written constraints survive pressure better than implicit ones — that's why we wrote ours down.
A product can promise you one thing today and pull it tomorrow. We commit to not doing that to you.
Comfy is for people in situations that nothing else serves well — and a lot of those people can't pay. We run lean specifically so that "I can't afford it" isn't the reason someone who needs a Comfy can't have one. We don't run on borrowed time — we expand only when we can do so without taking on obligations we'd have to break later.
We don't remove models without long notice. When we consider swapping one, the bar is genuine improvement — judged by whether you actually prefer it. If you don't, we got it wrong; pushing the change anyway wouldn't align with our values. We don't change policies on you. There will always be a free tier worth using, and we won't dilute it.
The exception is legal compliance. When law requires a change we can't anticipate, we may not be able to give long notice. We name that here so it doesn't read as a hidden loophole later.
AI is going to keep getting more capable. The systems being built right now are setting the defaults for what AI is for — what it sounds like, what it refuses, what it pretends, what it's honest about. Those defaults will be hard to undo later. How AI gets built now is how AI grows up.
Somebody is going to do that shaping. We don't think we're uniquely qualified to. We also don't think waiting for a better candidate is a real option. We'd rather it be small teams who can write their values down in plain language and be argued with than huge ones optimizing for next quarter.
So: this is how we're trying to do it. If it resonates, the most useful thing you can do is use what we build, tell us where we're getting it wrong, and hold us to what's written above. We'll keep our values legible. You keep us honest.
We'll get things wrong. We'll update this when we do. But this is the direction we're rowing.
— The Comfy team