MoaLLM Board

There are three ways to bring the thinking, and all three look the same on the canvas. MoaLLM's own plan opens a session with one click and no sign-up, as a free trial. An agent you already use connects over MCP. Or you paste an API key and the built-in agent answers the sessions you start.

Every tool is defined once: its name, the description the agent reads, its arguments and the call it makes. There are 60, from opening a session to writing an exam or reading a past paper. The built-in agent does not carry them all: a small model reads each message and loads only the groups it needs, a core of about 5k tokens instead of 14k.

Your vault is a folder of plain Markdown. You can write in it from the Board, with the note drawn as you type; an agent saves a note through a step that validates it, reindexes it and can be undone. Layout never goes into Markdown.

Recall is a record, not a schedule. The Overview counts the notes that survived an unaided retrieval, and an unaided grade has to carry a receipt: the sentence the learner actually said. Nothing counts down or keeps a streak.

The graph window draws the vault from that Markdown on each request. Shape says what a note is and lines say how notes link. What breaks the vault's rules is judged in red pen and named: an orphan nothing links to, a prerequisite not written yet, a link that only goes one way.

Map is the drawing. Path puts what to learn first at the top, each note below what it needs. Time lays dated notes on a ruler that ticks from centuries down to days: spans as bars, causes as curves to what they led to, Hijri years one switch away. Live is the notes in motion: dragging one pulls its neighbours like a spring web, and under reduced motion it holds still.

The Meaning view lays every note out by what it means, as search sees it: each note's passages are embedded, averaged, and brought down to a flat map (or a 3D one you can turn) with PCA and UMAP. Notes on the same idea sit together whatever language they are written in, and colour says which hub they belong to.

Bring in your YouTube watch history from Google Takeout and every video you watched joins the map, placed among the notes it is near. Videos near none of your notes gather in a ring around them, named by their most watched channel. Thousands show as soft regions, then dots, then YouTube's thumbnails as you zoom in. Point at one and lines go to its notes; keep one and it becomes a source note.

Ctrl+K searches what was said in those videos too, and a result opens the video at that second.

A good multilingual embedding model is slow on a laptop. EmbeddingGemma takes about 175 ms a query on the CPU, too slow for a search box that answers on every key. So the Board has its own: Qatra (قطرة, "a drop"; Arabic's word for distillation shares its root).

Qatra is a static model: a table of token vectors, averaged per text, so no neural network runs when you search. I distilled it with Model2Vec from two teachers, EmbeddingGemma 2 and IBM's granite-embedding-311m-multilingual, over a corpus I built from documentation, lecture transcripts, parallel Arabic and English pairs, and generated learner questions and notes in English and Egyptian Arabic. Then I reweighted it so common words count less, and quantised it to int8.

One model now does what two did before, one for English and one for Arabic, with no language router in front.

Tell your agent a date ("my final is on 21 December") or share a syllabus, and it proposes the dates. Each waits at the top of the Calendar until you tick it; only ticked dates count down, in the heads of the course's hubs and on the Overview. Your university's .ics comes in, and "Save as .ics" takes the dates to your phone.

Add a past paper and the agent splits it into its questions, each with its page, marks and kind. The Board checks each question on its page, matches it to a note on your computer, and from every paper of a course writes a pattern note: per topic, how many papers asked it and its share of the marks. Nothing reminds you, and nothing is due.

Sections are named frames that hold windows, like the parts of a map: drag one and its windows come along. Ask the agent to tidy the desk and it groups what belongs together.

Dock a session as a full-height panel on the right and read on the left; a window the agent opens lands beside the panel, not under it.

Looks change the whole Board: Air, Graphite, Earth, Tide and Summer, each light and dark, any colour theme on any look, and nine accents. Notes can also draw their Mermaid blocks in place, as clean shapes in the Board's colours.

An agent over MCP reaches the Board through a local discovery file, readable by you only, and a loopback server on a random port.

Every request carries an ID, so a retry returns the original result instead of a second message. Each canvas change carries the revision it was based on: if you dragged a window after the agent read the scene, the agent's move fails as stale instead of overwriting yours. The status line shows only states that are true.

1,147 automated tests, counted on 20 October 2026, run on every push on Ubuntu and macOS together with the installer. A browser scenario drives a real Chromium through an Arabic session, a marked question, a shared PDF page, a sheet written with a pen, the built-in agent streaming from a fake model server, and a server restart. The web suite also checks text contrast in every theme.