Mavéa：把答案画在画布上的 AI

Mavéa™

An AI that draws the answer.

Ask a question — type it or say it — and Mavéa builds the reply on a living canvas: charts, timelines, maps, and evidence you can check, with its pen circling the exact figure each line is about. See it in motion →

Talk to AI. See what it means. If that is a direction worth exploring, a ⭐ on GitHub helps other people find Mavéa.

npx @mavea/mavea@latest

Opens http://localhost:4173 — no install, no account, no model key required.

_{Runs locally · Your own API key · OpenAI, Claude, Gemini, Grok}

_{Ask "how do we plan a day trip?" — and the answer arrives as a route map, an hour-by-hour playbook and the totals that matter. Mavéa picks the form each answer needs.}

_{Ask "why did it happen?" — and the answer opens into the causal web behind it. Turn a cause down and the whole web re-weighs, with every link traceable to a real quote.}

Get started

All you need to run the published app is Node 22.12+; building Mavéa from this repo is stricter and wants Node 24.15+. That command above opens the tour and demo replays immediately — nothing else to set up. Demo replays are fictional, curated prerecorded examples with scripted feature choreography; playback does not call a model provider.

The @latest tag matters: for a bare package name npx can reuse a previously cached copy rather than fetching a newer one, so asking for latest is what reliably gets you the current release. Prefer a pinned local copy? npm install -g @mavea/mavea, then run mavea.

To talk to a real model (Live): click "Open Mavéa" and paste an Anthropic / OpenAI / Gemini / Grok key. Model usage is billed by that provider under your key, at its rates — the model call is the part of a turn that costs money; the app and the speech services run on your machine. The key stays in memory unless you opt into encrypted local remembering, and each provider request carries it through your same-origin proxy to that provider — your key and prompts never pass through Mavéa's own servers, because there aren't any. Settings → Your data → Forget everything on this device removes the key, the encryption key that sealed it, and everything else Mavéa stored in your browser. Full options — models, hosting, and the trust boundary — are in docs/LIVE-SETUP.md.

Local speech: Mavéa speaks through Apache-2.0 Kokoro weights and wrapper and transcribes through MIT-licensed whisper.cpp. Both cost nothing per use, and what they produce is yours: publishing a transcript, or a video spoken in Mavéa's voices, owes no fee and no credit line — those licenses cover the software, not its output. The defaults run on your machine through loopback-only proxies; a deployment that overrides WHISPER_URL sends microphone audio to that configured endpoint. Mavéa runs without the voice services, with the canvas, captions and typing intact, but you will not get spoken replies or local mic transcription. For the full experience, install a container runtime and let npx @mavea/mavea or pnpm dev start the two services. Podman is the recommended open-source runtime; Docker also works (Docker Desktop is licensed separately). Audio is never handed to a browser-vendor speech service as a fallback.

During a conversation, the Mavéa's voice toggle turns output speech off without changing the microphone. A paced answer then reveals in full immediately, with captions, notes, and Pen marks in place of the spoken walk.

What it does

Mavéa is an AI you watch think. Type a question or ask it out loud: a calm face takes it, then steps aside while a living canvas draws the answer in charts, timelines, and evidence you can check, choosing the right form for each one. As it narrates, it draws on the answer, circling the exact figure each line is about. Sourced claims keep their links and estimates can be labelled.

A few specific things worth trying:

• Highlight to ask. Drag across any part of the answer and ask about it — the next turn is grounded on that exact bar, row, or number, not the whole card.
• It draws while it talks. It speaks each headline once the first card is on the canvas, then lands hand-style circles and arrows on the figure it's narrating.
• One board, and a card you can look closer at. The board lays the whole canvas out at once — nothing staged, nothing to pick. Click any card and it opens on its own stage, centred over the blurred board, with Mavéa's notes beside it: what it assumes, the pattern she sees, what the sources actually back, and the question that would break it. Step through the rest with the arrows, magnify if you want to, and Escape puts you back. When you'd rather be walked through the whole thing, Guide me pulls the answer onto a lamplit desk.
• It maps your thinking. Ramble for a minute and Mavéa clusters your words into the themes that surface from what you actually said — never fixed buckets — with the tensions between them.
• The Blank Space. When an answer needs a number only you have, Mavéa leaves a hole to fill — by voice, type, or a dragged card — instead of quietly guessing.
• View any "why" as a living answer. The answer opens into the causal web behind it: press Walk me through it and the camera flies cause to cause while Mavéa narrates, or read the same web as contribution ribbons, spheres of force, a timeline, or a chart. Break a cause open and its parts are drawn by whichever component the library has for that shape — or simply named, where nothing measured them. Every figure on it can prove itself, every arrow says what it does not claim, and pulling a what-if lever re-weights the world in words — never in invented numbers.

The full tour — the Study, Prism and Synthesis, Deep Zoom, Courses, Ripple, the Atlas, the Rehearsal, living dashboards, selective Conversation video, Mavéa Reels, and deck/document export, among some thirty features — is in docs/FEATURES.md. In the app, press ⌘K.

What it looks like

Twelve moments from the walkthrough and the recorded sessions. Every one replays with no key via npx @mavea/mavea@latest — except the thought map, whose threads need a live model to draw.

It marks what it's saying. The pen lands on the exact line as the sentence is spoken.	Guide me. The answer comes forward a card at a time on a lamplit desk, Mavéa’s notes pencilled in the margin.	Think out loud first. A ramble sorts itself into themes — and it names the tension you missed.	Scrub the voice. Drag the spoken track and the canvas un-builds to what had been said.
Walk the why. A "why" answer opens into the causal web behind it — every cause traceable to a real quote.	It builds you a course. Ask to master something and get real lessons — objectives, a checkpoint, next steps.	A topic, in powers of ten. Telescope from the big picture to the finest mechanism.	A document, claim by claim. Every finding highlighted on the real page it came from.
It picks the right form. A route becomes a map; a day becomes an hour-by-hour plan.	A repository, taught. Beginner to expert, built from the code, with real files.	Out as a deck. Ten skins, a live preview, PDF or PowerPoint.	Or as a document. The same answer set as a paper, ready to print or send.

What it needs

Mavéa runs on your machine, not in a cloud.

Any browser in Baseline Widely Available — current Chrome, Edge, Safari, or Firefox — on macOS, Windows, or Linux. Minimum covers the app with voice off, alongside the browser and whatever else you already have open. Recommended adds the local Kokoro + whisper.cpp services and a container runtime, with enough room that you never think about it.

Local speech costs real memory, and it's optional. Kokoro holds roughly 1.3 GB resident; whisper.cpp adds its local model and working memory; a desktop container VM adds its own overhead. At the 8 GB minimum, leaving local speech off can be the better trade. podman compose down (or the Docker equivalent) reclaims it whenever you want it back.

Third-party commercial use

Generated media, bundled libraries, fonts, maps, and the local speech services are held to an open-media and permissive-licence allowlist, enforced on every build by pnpm check:licenses. It is an engineering control, not a legal opinion — the full policy, what the gate does and does not cover, and the obligations that travel with the speech container are in docs/THIRD-PARTY-POLICY.md.

📄 Mavéa license

Mavéa is source-available under the PolyForm Noncommercial License 1.0.0. It is not licensed under MIT and is not distributed under an Open Source Initiative-approved open-source license. The PolyForm license permits use, modification, and distribution only for purposes it defines as noncommercial. Commercial use requires separate written permission or a separate license from the applicable rights holder.

That restriction applies to recipients of the PolyForm license. It does not stop an applicable rights holder from selling or commercially licensing Mavéa, offering it under different terms, or transferring rights it owns or controls, including as part of an acquisition. Third-party components remain under their own licenses.

Copies of the software must keep the license's required notice:

Required Notice: Copyright (c) 2026 Akash Maitra and Aryan Chordia

Contact

• Anything else — trymavea@gmail.com
• Feature ideas and bugs — open an issue
• Security — follow SECURITY.md; never report a vulnerability in a public issue