LetMeCheckThatBot.
It turns the group chat
into the interface.

Each kind of artifact gets its own reader, and they all output the same thing: text. A pasted link is fetched and opened in a headless Chromiuma full web browser running invisibly, with no window, driven entirely by code., scrolled the whole way to wake the lazy-loaded parts, and captured top to bottom. If the link is a video, yt-dlpan open command-line tool that pulls a video’s title, description, and transcript. pulls its title, channel, description, and transcript. A screenshot or image is handed to a vision modelan AI model that can look at an image and read or describe what’s in it. that transcribes the words inside it verbatim. A video or silent GIF is sampled into frames and read across them. A shared file has its contents pulled in. A voice note nobody played goes to a local Whisperan open speech-to-text model that turns voice notes into text. container and comes back as a transcript. Once everything is text, one memory holds it all, and the bot can answer about a menu photo or a music video the same way it answers about a typed line.

This is why ingest happens the instant a thing lands. The question arrives weeks after the artifact. If the voice note was not transcribed on arrival, the answer is not there when someone asks what that restaurant was. Silent ingest is the whole memory.

The core is the primitive Penny uses and the one Anthropic argues for in Building Effective Agents: an augmented model in a loopthe whole architecture in five words: one AI model that can read, decide, and call tools in a repeating cycle until it has an answer.. The bot reads the recent thread, decides whether to act, and if it acts it calls a toola capability the model can invoke on its own: search the web, read a link, build a meme, look up a memory., reads the result, and decides again, up to four passes before it owes the chat an answer. There are nineteen: it searches the web, reads and screenshots links, digs through Reddit, finds a video and clips the part that matters, builds memes and gifs, reads images and the text inside them, transcribes voice notes with one local Whisper container and answers out loud through another, synthesizing its reply into a Telegram voice memo with a local Kokoroan open text-to-speech model; it generates the spoken voice locally on the laptop with no cloud call. model, recalls the chat’s own history by meaning, and can even put a quick either-or to the group.

One detail earned its place the hard way: not every model honors the function-callingthe standard way a model asks to use a tool: it emits the tool’s name and arguments in a fixed, machine-readable format. contract, so the loop carries a recovery path that parses malformed calls back into real ones instead of giving up. That single piece of defensiveness is the difference between a loop that works on one model and a loop that works on whatever model I drop underneath it.

The moment a message lands and the translators have turned it into text, the row is handed to a single background worker. The worker calls a local embedding model, Qwen3-Embedding-0.6B quantized to Q8shrinking a model’s numbers so it runs on modest hardware, trading a sliver of precision for speed and size., running on Ollamaa tool for running AI models locally on your own computer, with no cloud and no per-use bill. on the same ThinkPad, and gets back a list of 1,024 numbers: the message’s coordinates in meaning-space. That vectora list of numbers that captures a message’s meaning, so similar ideas land near each other. is written straight into the message’s own row in SQLitea tiny database that lives in a single file on disk, with no separate server to run., next to the text. There is no separate vector database and no embedding API. The worker takes one call at a time so it never crowds the GPU that Penny is also using, and the whole thing costs nothing per message.

A hosted embedding API would have quietly turned more than 92,000 messages into a recurring bill, which is most of why this runs locally. Cost is the smaller half. The bigger half is dependency: a model on my own shelf cannot raise its price, deprecate its weights, rate-limit me, or go down at the wrong moment.

Recall does not pick between keyword and meaning, it blends them. It embeds the question once, uses a binary-quantized sqlite-vec indexa vector-search extension inside the same SQLite file. It narrows the room to the most likely candidates without adding another database service. to retrieve candidates from that room, then re-scores those candidates against the full 1,024-dimension vectors before merging the strongest hits with exact keyword matches. So “that thai place” finds the right night whether you remember the words or only the feeling. The index stays inside the same SQLite file, so search is fast without adding another service. The full vectors remain the durable asset; the index is replaceable.

It also sketches each regular from their own messages, so it knows who is in the room before it speaks. The result is a multimodal RAGretrieval-augmented generation: fetch the few relevant memories first, then answer from them, instead of guessing. “Multimodal” means it works for images and audio too, beyond plain text. over the group chat: the room’s whole history, in every modality, instead of the last fifty lines.

The bot reaches models through OpenRouter. MiMo v2.5 remains primary because it tool-calls reliably, holds a large contextthe context window: how much text the model can hold in mind at once. A bigger one means it can read more of the room before answering., and stays within a dedicated key hard-capped at $5 a month. Gemini 2.5 Flash Lite is the hosted fallback for a provider or model failure. If both hosted paths fail, a small local Functionary model can answer from local knowledge with a narrowed tool set. This is not the old budget cascade that quietly traded quality for price. The fallbacks protect availability, not a silent downgrade.

A $5 cap invites an obvious question: is a model this cheap any good? So I measured it. I gave the same model real work from GDPval, OpenAI’s set of actual job tasks across 44 occupations, each with a human professional’s own finished work and an expert grading checklist. Then Claude Sonnet 4.6, a model from a different company, scored its work and the human’s the same way. I ran one task from each of 44 jobs. The standard setup averaged 76, this bot’s harness averaged 70, and the human experts averaged 84. It also failed quietly: asked to build a musician-payroll spreadsheet it produced a flawless one with every pay rate off by about a dollar, so every paycheck was wrong. The failures look like finished work. The narrow, real thing it proves is that the floor for this kind of work is higher, and far cheaper, than it looks. Every task, output, and score is published, so you can judge it yourself in the benchmark explorer.

A bot that pings on a schedule is noise. So the trigger is a dying room: a full week of real silence, never the calendar. The payload is the room’s own best day, this same day in an earlier year, posted back as a flashback in everyone’s own words. The notification is suppressed, so it surfaces only when someone next opens the chat. You find it; it never pings you.