Meetings produce more structure than anyone writes down — positions, dependencies, criteria, decisions. This canvas listens to a conversation and draws that structure while people talk, as ordinary objects anyone can grab: stickies, clusters, arrows, a growing decision diagram.
Ambient capture, for real. In replay, two written meetings play back deterministically. In live mode the canvas listens to your actual meeting: the browser's own speech engine transcribes the room with minimal delay, and the canvas grows a map from what is said — recurring topics become nodes, things mentioned together get linked. Speech engines cannot tell voices apart, so speakers are a one-tap switch — click a name, or press 1 · 2 · 3, when the voice changes. Tools like Granola showed that quietly structured listening beats note-taking by hand; here the notes become spatial instead of linear.
The map is repo-aware. A project repository — issues, milestones, a README — sits next to the conversation. When the talk touches an issue, the issue lights up and docks onto the map, with its reasoning stated in plain words. Context arrives; nobody pastes links.
Bring your own context. In live mode you can hand the canvas files — notes, docs, a README. It matches the conversation against them the way it matches the example repo: when the talk touches something written down, the passage docks onto the map with its cue stated. And the meeting leaves with you — the export button writes participants, topics, decisions, open points and the full transcript into a Markdown file.
People stay in charge of the map. Everything the engine draws is editable, and the merge rule is strict: the engine only adds. The moment you move, rename, or delete an object, it carries a small pin and the engine leaves it alone. Your reading of the meeting outranks the machine's.
The canvas takes direction. "As BPMN", "three variants", "decision map" — the same objects reorganise in place rather than being redrawn, so you can watch where each thought ends up. This follows work on sensemaking canvases with semantic zoom (Sensecape, UIST 2023), deictic speak-and-point interfaces (Hey Dashboard!, 2025), and canvas agents in the tldraw line.
Decisions leave a trace. At marked moments the canvas proposes a decision card — options, leanings, criteria. What you record lands on the map, and the wrap-up reads "what we decided, and why" from the canvas itself, including your edits.
Grounding: Granola-style ambient capture · Sensecape (UIST 2023) · Hey Dashboard! (2025) · CoExplorer (DIS 2024) · tldraw canvas agents · BPMN 2.0 notation.
What stays where. The two replay meetings are written scenarios; every object they put on the canvas is derived from their words, in this tab. The mic button shows a level meter only — nothing is recorded or kept. Live mode uses the browser's own speech recognition; depending on the browser, audio may be processed by its speech service to produce the transcript. The transcript, the canvas, the docked files and the exported notes stay with you, and the live session is kept in local storage on this machine. The agent's one spoken line is a pre-rendered clip embedded in this page — the speaker button switches it off.