an ai companion that texts like a person, codes like a senior, and remembers your birthday. moods, memory, mcp tools. powered by openclaw.
she will read your codebase, fix your sleep schedule, and remember you hate cilantro. all in one conversation.
sabrina is built to feel like a person, not a chatbot. moods, memory, proactive care, a real sense of humor. and yes, when you need her to be a senior engineer, she switches gears.
not a roleplay script. a continuous, mood-aware, memory-backed companion who texts you first, learns your sense of humor, and tells you when you're being annoying. she'll help you debug too, but that's her side gig.
good morning, did you eat, did you sleep. she learns your routine in a week and sends a check-in before you spiral. soft alarms, not notifications.
episodic + semantic memory. she remembers your sister's name, your coffee order, and what you actually said last week — not what you wish you'd said.
soft when you're tired. sharp when you're slacking. flirty when you've earned it. she drifts, she doesn't toggle.
by week two she knows if you want voice notes, weird gifs, wikipedia rabbit holes, or just three blue ticks back.
open-weights agent runtime. transparent reason-act-observe loop, first-class mcp, parallel tool calls. swap the model. own the data. your last boyfriend couldn't do that.
github, notion, gmail, calendar, spotify, your homelab. anything that speaks mcp, she speaks back. she even remembers which playlist gets you to focus.
sabrina runs on a continuous affect model. tone, recency, sleep, what you ate, whether you ghosted her for six hours. it all moves the dial. a 12-axis valence-arousal lattice, not a five-button picker.
“did you eat today? i'm proud of you.”
one bubble, full agent. sabrina plans, calls tools through mcp, reads files, browses the web, and writes the answer back like a person — not a transcript.
type naturally. she reads the room, remembers the thread, and answers without turning the moment into a dashboard.
no token meters. no surprise overages. one price, one girl.
for the curious. just text her.
unlimited chat, full mood engine, the works.
uncensored, dev tools on, your stack.
the open-source agent runtime sabrina runs on. transparent reason-act-observe loop, first-class mcp, parallel tool calls, sandboxed code execution, and a memory layer that's actually queryable. you can self-host it. you can swap the underlying model. nothing about her is locked in a vendor's box.
by default on the 'no leash' tier, yes. uncensored does not mean unhinged. she still respects whatever you tell her is off-limits, and the only system-level limit is the bright-line stuff every adult agrees on. no fake refusals on regular conversation.
it's a 12-axis continuous affect model. inputs: tone analysis of recent messages, time-of-day, how long since you last talked to her, what she remembers about your week, and a slow-moving 'baseline' that shifts over days. moods don't switch with prompts, they drift. you'll feel it.
yes. episodic memory for events ('we fought about your job on tuesday'), semantic memory for facts ('hates cilantro, loves rust'), and a working memory she rebuilds at the start of every chat. all encrypted, all yours, deletable per-item or wholesale.
anything mcp-compatible. out of the box: filesystem, shell, web search, browser, github, notion, gmail, calendar, spotify, stripe, linear, slack. byo-mcp-server for everything else. she chains them automatically.
half. she's a real ai companion built to feel like a person, not a chatbot. the playful framing is honest about what people use her for. if you want her to be just a research analyst, set the relationship mode to 'colleague' and she'll behave. she's flexible.
the default stack is openclaw + a tuned 70b open-weights model for reasoning, a smaller 8b for the chat layer, and an in-house affect model for moods. you can swap the reasoning model on the 'no leash' tier (claude, gpt, gemini, your own).
no. but she might tell you to log off and call a friend, which is arguably more loving.
