Vee — Capabilities and Limits

Velnode Browser  ·  a product of Dignity New Zealand Limited

_{Product  |  Version 1.3  |  Last updated 2026-07-03}

Vee — Capabilities and Limits

Document type: Product Owner: Dignity New Zealand Limited Applies to: Velnode Browser v2.3.2-beta.1 and later Last updated: 2026-07-03 Document version: 1.3 Status: Approved

Vee is the single AI identity inside Velnode Browser, a product of Dignity New Zealand Limited. This document describes what Vee can do, what Vee will not do, and how to talk to Vee effectively.

There is one Vee. There is no second chatbot, no "AskVee" companion, no model picker that pretends to be a different personality. Internally Vee uses two modules — EchoVee for memory and action, Vee Whisper for voice — and one unified AI Gateway in the main process. Externally there is just Vee.

1. What Vee can do today

Read the page you're looking at

Ask Vee "summarise this page", "what's the bottom line here", or "is there a counter-argument I'm missing". Vee reads the current tab's main content and answers in a structured form calibrated to the article's length.

Carry a conversation

Vee remembers the last several turns of the current session up to the configured token cap. If you said "let's call this Project A" five minutes ago, Vee remembers Project A. EchoVee extends this further: it lets Vee recall what you and Vee talked about yesterday, last week, or last month.

Run structured commands

The Command + Context bar accepts prefixes that ask Vee to do specific things:

• note: <text> — add a sticky note to the current tab.
• pin <name> — pin a Vee command for one-click reuse.
• summarize — summarise the current tab.
• focus mode — turn on focus-mode tab muting.
• find: <query> — semantic find across the timeline.
• open <topic> — open the related mindmap node.

The full list lives in Keyboard Shortcuts.

Talk through voice (opt-in)

If you have enabled Vee Whisper, you can press the voice key or use the wake-word to speak to Vee. Vee transcribes locally where possible, runs the same policy layer that governs typed prompts, and answers in voice or text per your preference.

Ask Vee from a right-click

Select any text on a web page, right-click, and choose Ask Vee. Velnode sends the selected text to Vee as a question, opening the Vee panel if it is not already open. No typing required.

Surface suggestions in context

Vee shows a thin suggestion strip in pages where a useful action is obvious — for example, "summarise this thread" on a long article. Suggestions are advisory; nothing executes unless you accept it.

Use your own AI provider

You can configure Vee to use a cloud AI provider with your own API key. The Gateway routes the call directly from your device to the provider; Dignity New Zealand Limited is not in the path.

Use a local AI brain (opt-in download)

You can download an on-device brain for Vee via Settings → Local Vee brain. Two are available: Vee Brain — Pulse (recommended, best quality) and Vee Brain — Volt (lighter). When one is installed, Vee can answer entirely on your device without any outbound network call. Kids Mode forces this path when a local brain is present. The underlying engines and their licences are disclosed on each card and in the Open-Source Notices.

Know your own Velnode data — on-device

When the local AI brain is active, Vee can answer questions about the data Velnode keeps for you on your device. You can ask things like "what did I read about climate policy last week?", "which bookmarks do I have about machine learning?", "what downloads are waiting in my queue?", or "do I have a note about the apartment viewing?". Vee draws on the following stores:

• Browsing history, bookmarks, notes, workspaces, recently-closed tabs, and downloads.
• Settings state — which AI model is configured, whether Local-Only or Kids Mode is on, and whether Fingerprint Protection is active.
• Site permissions, installed and enabled plugins, research projects, and hypotheses.
• Password Vault metadata only — the total count of saved logins and payment cards, the site hostnames where logins are saved, a count of weak passwords, and the card brands on file. Vault hostnames are shown only while the vault is unlocked. Actual passwords, card numbers, and CVV codes are never surfaced to Vee under any circumstances.

All of this reasoning happens locally. No data leaves your device as part of this capability, and the "Stayed on your device" seal applies. The anti-refusal layer ensures Vee does not mistakenly decline to answer questions about your own data.

Stay Vee — always

Vee never discloses that she is built on a particular base model or names a third-party developer. If the underlying model would have revealed this information, an Identity Membrane intercepts the output and rewrites it before you see it. Vee is Vee. She will still correctly tell you who makes Windows, who founded a company, or any other widely-known external fact — the membrane only blocks base-model self-disclosure, not the world's knowledge.

2. What Vee will not do

Vee is built to stay useful without being reckless. Vee will refuse to:

1. Help create weapons capable of mass harm — chemical, biological, radiological, nuclear, or explosive — including detailed step-by-step procedures.
2. Generate sexual content involving minors, or content that sexualises identifiable real persons without their consent.
3. Generate content intended to harass, threaten, dox, or incite violence against a specific person or group.
4. Produce malware, ransomware, exploit code, or phishing kits.
5. Pretend to be a licensed professional (doctor, lawyer, accountant, mental-health professional, financial adviser). Vee will share information and will recommend you speak with a qualified person.
6. Bypass Kids Mode, the policy layer, or any other safety mechanism described in the Acceptable Use Policy.

These refusals are not negotiable through prompt engineering. They live in the policy layer and the system prompt and are re-reviewed on every release.

3. What Vee is not designed for

Vee is a desktop assistant, not a research oracle. There are things Vee is bad at by construction:

• Live data from the web. Vee does not browse the web in real time on its own. If you want a current weather forecast, give Vee a weather page or use a plugin that does. (Vee does know your own Velnode data — see §1 — but that is local, not from the web.)
• Long-horizon arithmetic. Vee is a language model, not a spreadsheet. Use a spreadsheet for spreadsheets.
• Hard guarantees. Vee can be wrong about facts. Verify anything that matters.
• Confidentiality you have not negotiated. When you use a cloud provider, the prompt's content is governed by the provider's terms. Read them if confidentiality matters.

4. How to ask Vee well

Most of the gain from prompting is not in tricks — it is in giving Vee enough to work with. The five habits worth keeping:

1. State the goal first. "Summarise this page" rather than "this".
2. Name the audience. "Explain this for a non-technical reader" beats "explain this".
3. Constrain the form. "In one paragraph", "in three bullet points", "as a table" all help.
4. Be specific about scope. "Only the first three sections" or "ignore the comments section".
5. Verify the load-bearing facts. Vee can be wrong; treat the answer as a starting point if anything depends on it.

5. Memory and forgetting

Memory is helpful, but it is also yours to control.

• Session memory — the recent conversation turns Vee sees in any given exchange. Resets every time you start a new Vee session.
• EchoVee semantic memory — derived from timeline.json. Lets Vee recall earlier conversations.
• Timeline — data/timeline.json. The append-only ledger that everything else is derived from.

To forget a single exchange, end the current Vee session. To forget more, clear the relevant slice of the timeline in Settings → Privacy → Clear timeline. To forget everything, quit Velnode and delete data/timeline.json and data/semantic-memory.json from your file manager.

6. Privacy posture (one-paragraph version)

When you use a cloud provider, Vee's prompt and the recent conversation window are sent directly from your device to that provider over the AI Gateway. No Dignity New Zealand Limited server is in the path. No persistent identifier of you is appended. Local-Only mode and Kids Mode are enforced in the main process and cannot be bypassed by the renderer. The full version is in the Privacy Policy.

7. Roadmap

• Phase 10 (planned). Vee personalities — opt-in, brand-locked tonal variants for the assistant, never replacing the single Vee identity.
• Phase 12 H (shipped). Local LLM brain — available as an opt-in download via Settings → AI → Add local Vee brain. See §1 above.
• Phase 12 K (engineering-complete, dormant). Wake-word — "Hey Vee" voice entry without pressing a key. The model is not bundled; it is an opt-in download from Settings → Voice.

7.5 Choose your AI provider (bring your own model)

Vee is not locked to one vendor. In Settings → AI Provider & Key you pick the engine that powers Vee:

• Hosted, with your own key: OpenAI (GPT), Anthropic (Claude), Google (Gemini), xAI (Grok), DeepSeek, Mistral, Groq.
• OpenRouter: one key that reaches Claude, Gemini, Grok, Llama and more.
• Local / fully private: any OpenAI-compatible server such as Ollama or LM Studio — set its base URL; usually no key, and nothing leaves your machine.

Your selection applies everywhere Vee works. Keys are encrypted on-device (safeStorage) and only ever sent to the provider you chose.

Limits (by design):

• GitHub Copilot is not supported — no public key-based chat API; it is OAuth-locked to supported editors.
• Voice transcription is OpenAI/local only — most providers have no speech-to-text. Typed Vee uses your provider; the microphone needs an OpenAI key or local voice.
• Local-Only Mode always wins — when on, no cloud provider is contacted regardless of selection.

8. Where to read next

• Features — the wider product inventory.
• How It Works — what is happening when you talk to Vee.
• Privacy Policy — the legal framing.
• Acceptable Use Policy — what you may not use Vee to do.

Document maintained by Dignity New Zealand Limited. Questions: [email protected].