Nothing open. Drop a SillyTavern card into the roster on the left, then hit chat, or 💬 if you'd rather she texted you.
Or don't.
Using your preset, your jailbreak, your persona,
and whatever's loaded.
Nothing open. Drop a SillyTavern card into the roster on the left, then hit chat, or 💬 if you'd rather she texted you.
Or don't.
Using your preset, your jailbreak, your persona,
and whatever's loaded.
Local servers are probed automatically. Remote keys live in data/config.json and never reach the browser.
Heads up: remote providers see your text. Your pictures stay on this machine unless you tick "send images" on a backend — meant for your own LM Studio on another box, not for a stranger's API.
CoomKit listens on this machine only, so a browser on your phone cannot see it however hard it tries. Turn this on and it listens on every interface: anything on the same wifi can open it — your phone, a tablet, another CoomKit pulling a datapack. There is no password on this app. Only do it on a network you trust, and remember that "trust" means everyone else on it too. Takes effect on the next restart.
A one-shot clone over your LAN: characters, chats,
memories, the gallery — everything. Your local backend settings
stay yours. The other install needs the switch above ticked before
it is reachable at all, and your current data survives as
data.bak/.
Normally her unprompted texts are scheduled by this
browser tab, which on a phone is suspended the moment the screen
blanks. Turn this on and the server keeps her schedule
instead — she can text while the app is closed, using the backend
and model picked right now. Which threads she may text at all is
still the bell toggle on each phone thread. On Termux, run
termux-wake-lock or Android will eventually kill the
server too.
A preset bundles a request mode, prompt template, thinking behaviour, sampling and a jailbreak.
Sampling lives in the scene rail: one set of sliders, not three. Tune it there and hit save into preset.
System-level framing. On local backends the reasoning prefill in a preset is usually stronger; remote providers strip it, so system text is what you have.
World info: facts about a setting that get injected when you mention them. A book attaches to a character, to one chat, or to everything — several at once, none of them merged into her card.
Import a SillyTavern World Info file, or a card that has a lorebook baked into it. I'll tell you what I can't honour before you commit, not after.
Find/replace rules, run in order. prompt rules change what the model reads; display rules change what you read. Both can be on. Neither ever rewrites the stored message. Your log stays canonical, exactly like macros do.
Patterns are SillyTavern-style JS literals: /foo/gi —
converted to Python. A rule that can't convert is kept and switched off with
the reason attached, rather than quietly never firing.
Your box, your workflows.
Which graph she reaches for, per kind of thing she is making. A character can override the image one on her own card; this is the fallback for everyone who has not.
Fourteen graphs already ship with CoomKit: anima, krea2, klein, z-image, H3, Wan, OmniVoice, MiniMax Music. They drive the studio's one-click shots and everything she asks for herself; you don't have to add anything to make her work. They're marked shipped below and are read-only, because they're files on disk that get rewired per render.
This box is for bringing your own. Export from ComfyUI with
Save (API format) and paste it here. Put
{{prompt}}, {{negative}}, {{seed}},
{{width}}, {{height}}, {{image}},
{{audio_text}} or {{music_prompt}} wherever you want values injected.
Your whole prompt, in order. Every block can be toggled, re-roled and moved, including past the chat history, which is where a reminder has the most force. CoomKit's own layers are in the same list as yours; nothing is hidden.
Every instruction CoomKit injects on your behalf. These are the invisible sentences that steer the model: director framing, texting rules, how she reasons in character, how the forge pitches scenes, what the memory extractor keeps. Edit any of them; blank or unchanged resets to the shipped default. Use the inspect button on the composer to see the result in context.
Presets and jailbreaks worth starting from, with honest notes on which lever moves which backend. Installing upserts by name, so your own entries are never touched, and re-installing just refreshes the shipped ones.
SillyTavern v2/v3 cards. Import a PNG or JSON; export re-embeds the card back into the PNG so it still works in ST. Edits here are written into the embedded card too, so they survive the round trip.
The one on her roster tile and every bubble she speaks in.
Used by every studio shot. Pin a seed and she stays the same person across the whole gallery.
Read live from your ComfyUI's
models/loras whatever that install reports is
what you see here. Stacked in order, using core nodes so they
work anywhere.
Anima's strongest style lever. one artist tag reshapes the whole image. Krea 2, Klein and Z-Image are natural-language models and these actively hurt them, so they're ignored there.
Cloning needs 3–15 seconds of someone talking. Use one of the shipped voices or drop in your own clip. yours always wins. Listen to it before you commit; a clone sounds exactly like whatever you feed it.
Paste a link and the seconds you want. It grabs
just that span, then checks it before keeping it — a
reference under 185 Hz can clone an octave down and come back
male, and nothing downstream tells you. Needs
yt-dlp and ffmpeg on the machine
running CoomKit.
Describe what you want, or nothing at all, and she gets invented around (you): your persona, and what you put in what you're into. Argue with the pitches until one is right, then commit and she gets a card, a pinned seed, a voice and a portrait.
You already found the picture. Card for that feel: drop it here and a vision model reads her off it — build, hair, what she's wearing, what the shot is actually about — then pitches several women who all look like that and are otherwise completely different people. The picture decides how she looks; the pitches decide who she is. She keeps it as her face and as her generation reference.
The picture never leaves this machine, so this needs a local vision model — a remote backend is refused rather than quietly shown nothing, unless you've ticked "send images" on it in ⚙ → backends because it's really yours.
One first_mes gets old fast. Brainstorm a
fresh situation instead, so she can build on what you two have already
done, or start clean.
Your own assistant. She is an ordinary character card underneath — give her a name, a face, a voice, a personality — except that she has no opening line, she stays out of your roster, and the assistant button always comes back to her.
Who you are in the scene. Swap per chat from the left rail.
For the recipes that involve you. On those, the video model gets your reference as <Picture 1> and her as <Picture 2> that's how you end up in frame instead of a stranger, and that order is deliberate: the first slot is weighted hardest and anatomy is what these models invent worst. Sent only for a recipe that asks for it (the act shots), only to your own ComfyUI, and never anywhere else. They live on your disk like everything else.
Every piece of this, and what put it there. Turn one off or edit its text and the prompt below changes.
Screenshots are for people who don't have me. Pick what goes in and I'll draw it properly: same bubbles you're looking at, no browser chrome, no tab bar, no clock.
how /g/ sees it in a reply
—