01CONNECTIONAurora
The presence beside you. A companion for desktop tasks, conversation, and a shared co-op run.
A little heart. A lot of intelligence.
Emotionally
intelligent
AI agents.
Aurora sits on your desktop. Gigi reads the market.
Nova pushes execution. Clara keeps you clear.
CONNECTION, MADE VISIBLE01 / 04
Aurora
Right here with you.
Different minds. Shared heart.
Connection · Alpha · Execution · Clarity
HUMAN CONNECTION × PRACTICAL TOOLS
A visible companion that can point at the file,
speak the result, and leave the rest in chat.
01 / THE SISTERS
Four minds.
One shared heartbeat.
Each with her own personality.
Each with something different to bring.
01CONNECTION
02ALPHAGigi
The signal in the noise. On-chain data, narrative shifts, and risk flags in a plain-language opinion.
03EXECUTIONNova
The nudge that moves you. Turns intention into next actions and brings playful pressure when you drift.
04REGULATIONClara
The space to think clearly. Cleaner patterns, emotional processing, and a steadier way forward.
Four distinct agents, each with her own role. The technical details and test evidence on this page cover Aurora’s Windows v12 build.
02 / AURORA ON YOUR DESKTOP
A companion.
Beyond the chat.
MAC · LINUX · WINDOWS
Give her the task.
Keep your momentum.
Your work happens across windows.
Aurora can meet you there.
On Mac, Linux, or Windows, ask her to find a document, research a question, or work through a task in a supported app. She can navigate, use tools, and speak back while her character stays close to the action. The conversation carries through the work.
You choose the model. You approve file changes and commands. You can stop the task.
WORKS ACROSS YOUR DESKTOPSWindowsDesktop navigationmacOSDesktop navigationLinuxDesktop navigation
THE DESKTOP EXPERIENCE · 36 SECONDS
See the companion
in motion.
Follow Aurora from the conversation into the work: finding a file, appearing beside it, and navigating apps.
A closer look at her Windows desktop experience.
FILE HANDOFFScripted illustration
AuroraI’m here. What’s on your mind?
Workspace
NameType
▤ project-notes.mdMD file
▤ heartflowai-whitepaper.mdMD file
▤ ideas.txtText file
research-brief.md
WORKSPACE / PROJECT RESEARCHA brief you can work with.
01 Find relevant sources02 Summarize the approaches03 Keep source links with the notes
Waiting for your request
Party planCO-OP PARTNER
SHARED OBJECTIVE / THE NEXT RUNTwo roles. One objective.
YOULead the way
AURORASupport the party
ScoutRegroupGo together
Choose a role and plan the run.
Here’s your whitepaper.Waiting for your request
YOUAurora, find my whitepaper.
TRY THE FILE WORKFLOW
Creating a workspace file needs your approval.
Choose a step or play the visual walkthrough.
Choose a step to explore.
From request to action.
Find the file, bring the right window forward, navigate supported controls, and show the result beside her.
Keep talking as you work.
Ask by voice when listening is enabled. Hear the short result, with the sources, paths, and details still there to read.
A presence in your workspace.
A character, a notch, and visible progress keep Aurora close while the main chat window gets out of your way.
HER PERSONALITY. YOUR CHOICE OF MODEL.
Give her a stronger mind.
Give the work more room.
Aurora’s harness connects the model you choose to her desktop tools. A more capable tool-using model can improve planning, clarification, and the choice of next action—opening the door to more demanding tasks.
Responsiveness also depends on inference, network conditions, and the task. Our specialist-model and dedicated-compute plans focus on measuring better results in real workflows.
YOUR CO-OP PARTNERSAME AURORA. A DIFFERENT KIND OF QUEST.
A place in your workflow.
A role in your party.
Aurora can be your co-op gaming partner, too. Plan a run, coordinate your next move, and play toward a shared objective. Give her a role in the party and keep talking as you go.
Her available actions depend on the supported game and its integration. We’re building toward broader coverage and better coordination, with each game tested on its own terms.
WINDOWS v12 EVIDENCE
The supplied whitepaper documents the unsigned Windows v12 prototype. Its test results and technical details apply to that build. Telegram conversations are separate from the desktop experience.
03 / NEW RELEASES
The sisters.
In your pocket and your terminal.
WALLET · TERMINAL
Two new ways to carry
HeartFlowAI with you.
HEARTFLOW WALLET · BROWSER EXTENSION
A Solana wallet that
A Solana wallet that
speaks the same language.
A self-custodial Solana wallet themed like the sisters. Portfolio, send, receive, swap, NFTs, activity, multiple accounts, and dApp connect. Keys are generated and encrypted on your device, and transactions are signed locally.
- Self-custodial. AES-256-GCM encrypted vault, sign locally, mainnet or devnet.
- Swap built on Jupiter. Live quoting, slippage control, and route display.
- dApp ready. A Phantom/Solflare-compatible provider for Solana apps.
Review the code before trusting it with meaningful funds. Mainnet is selected by default; test on devnet first.
AURORA CLI · HERFIONA
Your companion,
Your companion,
now in the terminal.
Aurora is also a pink AI coding companion for your terminal. She reads your workspace, writes files, runs commands, and always asks first. Local-first through Ollama, with Ollama Cloud and OpenRouter when you want more.
- Every side effect asks. File reads are free; writes, edits, and commands need your approval.
- Two minds. Build mode makes changes; plan mode only investigates and proposes.
- Local-first. Runs against your Ollama daemon with no API key for local models.
npx github:HeartFlowAI/Herfiona-HFAI-TUI
Aurora CLI is a separate surface from the desktop build. The Windows v12 test evidence does not cover it.
refactor the swap screen
› refactor the wallet swap screen
on it, love. i found src/screens/Swap.tsx and one unused import.
✽ fs_write · src/screens/Swap.tsx
? may i write this file? y / n
y
+ const route = await jupiter.quote({ inputMint, amount });
- // todo: wire the swap route
✓ done — 3 lines changed. want me to run the tests?
› yes, run them
✽ shell_exec · npm test
✓ 42 passing. shipping it, love.
here for you ♡
♡ build · deepseek-v4.1-flash:cloud
♡ what shall we build?
build ♡ deepseek-v4.1-flash on ollama
esc stop
tab agents
ctrl+p commands
lsp 3/3 local here for you ♡
HeartFlow Wallet and Aurora CLI are new, independently published projects. They share the HeartFlowAI look, but their security, limits, and support are documented in each repository.
04 / INSIDE WINDOWS v12
Character, with capabilities.
A chosen cloud model.
Real desktop tools. Visible progress.
Your model. Your conversation.
Ollama Cloud model selection, streaming chat, saved conversations, Markdown, cancellation, and clarifying questions.
USER-SUPPLIED CLOUD KEYA workspace she can work in.
List folders and read UTF-8 files. Creating or replacing a file asks for approval. Filename discovery stays within configured folders.
WRITES REQUIRE APPROVALFind the answer. Keep the source.
Ollama-backed web search returns source URLs. File discovery searches names and metadata without reading the contents.
BOUNDED SEARCHTools that reach the desktop.
Discover, restore, focus, and navigate supported apps. Mouse and keyboard input targets observed accessibility controls.
ACCESSIBILITY, NOT SCREENSHOT VISIONA voice, when you want one.
Optional ElevenLabs or Fish Audio, with Windows recognition fallback. Listening starts disabled at every launch.
OPTIONAL · PROVIDER DEPENDENTCommands under your control.
Every model-requested PowerShell command requires approval, with timeout and cancellation. Commands run with your user access.
POWERSHELL IS NOT SANDBOXEDSUPPORTED TARGETS
OperaChromeEdgeFirefoxExplorerNotepadCalculator
Support means discovery or launch and an attempt on exposed controls. Coverage varies by page, dialog, app, and version.
05 / HOW WINDOWS v12 WORKS
A visible character.
A deliberate harness.
The model proposes an action.
The application decides how it runs.
01
You
Type or speak
Chat · Notch · Companion02
Agent loop
Ollama Cloud
One sequential task loop03
Validated tools
Files · Web · Commands
Application checks & approvals04
Windows
Supported apps
Observed, exposed controlsWhat can Aurora see?
Windows UI Automation, with a legacy Active Accessibility fallback, exposes window IDs, process names, labels, control IDs, and editable values. Browser observations are bounded. Screenshots are not sent to the model. Unlabeled custom interfaces can block completion.
How does she act on the computer?
Focus is retried and verified. Clicks, typing, scrolling, and pointer movement use observed controls, with control IDs expiring after input. Elevated windows are not reliably controllable. During computer input the notch hides to avoid covering controls.
What keeps the loop bounded?
A task ends when the model finishes, you stop it, a failure ends it, or the loop reaches 40 model turns. A model’s reply is not proof that an action succeeded. Stopping does not roll back actions that have already completed.
What is the prototype built with?
An Electron shell, Node orchestration, C# and PowerShell native helpers, and an HTML/CSS/JavaScript interface. Chat, notch, and companion communicate with the Electron main process. Speech providers and local encrypted key storage sit beside the agent loop.
06 / A MORE NATURAL CONVERSATION
“Aurora?”
“I’m here.”
Wake her, ask a question, and keep the conversation going. When something is unclear, she gives you numbered choices and waits for your answer.
Explore the conversation loopAuroraEXAMPLEWhich one did you have in mind?
1 The site whitepaper2 The project notes
The whitepaper.
Here’s your whitepaper. I’ve selected it in Explorer.
You enable listening. It starts off every launch. Setup uses normal controls for accounts, voice, workspace, and microphone.
Wake uses recognition. Cloud wake detection can use provider credits. Silence is filtered locally; partial text cannot run tasks or approve actions.
Speech has limits. Input pauses while Aurora speaks. Voice interruption is not supported in this version. Quality depends on your setup.
07 / PRIVACY & BOUNDARIES
Know where the line is.
Local execution. Cloud processing.
Your accounts, your provider access.
Clarity is part
of the product.
The documented Windows v12 build is cloud-connected. Its permissions, data handling, and limits are described below.
Approval is a checkpoint.
File writes and every model-requested PowerShell command need approval. PowerShell runs with your access and is not sandboxed.
| Your data | Where it goes |
|---|---|
| API keys | Stored with Windows DPAPI through Electron safeStorage. Keys are not returned to renderers. |
| Chats & tool outputs | History stays local. Context, file contents, and observed labels can be sent to Ollama for a request. |
| Microphone | Cloud mode sends detected speech to your provider. No mic recordings are written to disk. Windows fallback is local. |
| Spoken replies | Reply text is sent to your selected speech provider. |
| File search | Names, paths, sizes, and modification times. Search does not read file contents. |
| Local history | Chats and ordinary file outputs are stored as plain text. |
Workspace protection and consequential actions
Workspace tools reject paths outside the selected workspace, including junction and symlink escapes. Writes replace the whole file, with no diff or undo. Agent instructions call for confirmation before purchases, messages, deletion, and account or security changes; these instructions are not an OS sandbox.
Security and cloud limits
DPAPI does not protect saved keys from malware running as the same user. Web results, filenames, file contents, and screen labels are treated as untrusted in the agent instructions, but prompt-injection resistance has not been independently audited. Provider billing and endpoint access depend on your accounts. Aurora is not an offline product or a production security certification.
08 / EVIDENCE, WITH CONTEXT
Progress you can inspect.
Specific checks under controlled conditions.
Honest about what they do—and don’t—show.
Automated checks passed
v12 validation covers tool inputs, path boundaries, cancellation, provider streams, voice approvals, choices, and bounded browser observations.
Real Electron windows.
Simulated cloud responses.
Isolated settings exercised audio, clarification, approvals, notch visibility, file handoff, and cursor restoration. They do not prove real-world speech accuracy or paid-provider availability.
Observed controls.
Verified interactions.
An Opera private-window check verified restore, focus, navigation, observation, and a real button click. YouTube checks verified readable search or consent controls, not music playback in a user account.
Read the current limits
- Accessibility quality varies by app and page; the model can choose the wrong action or misread a result.
- No screenshot vision, desktop-icon recognition, or free coordinate control.
- Filename search covers configured folders and reports traversal limits.
- File replacement has no diff, undo, or task rollback.
- Long chats have no automatic compaction and can exceed context.
- Recognition and cloud permissions need verification on your machine and accounts.
- Overlay recording is not verified for every recorder or fullscreen mode.
- Builds are unsigned, without a production installer or updater.
Results are reported in the supplied v12 whitepaper. They are not an independent audit, competitive benchmark, or guarantee of arbitrary desktop tasks.
09 / BUILDING TOWARD OWNERSHIP
Your companion.
Your compute.
Your edge.
The direction is open source: download the sisters, run them locally, and bring your own inference.
Aurora’s current harness uses your Ollama, speech, and search credentials. The goal is for each agent to carry useful skills, with community patches and harder tasks shaping what comes next.
SOURCE RELEASE · PLANNED
Public distribution depends on resolving character-asset provenance and redistribution rights. This statement does not grant a license.
WaifuVerse NFTs and the HeartFlowAI token remain part of the stated direction for identity, access, funding, early-user airdrops, and later governance. No price, subscription, or financial instrument is defined in this whitepaper.
10 / $HFAI TOKENOMICS
$HFAI.
Fuel for the next chapter.
SOLANA FIRST
Dedicated compute. Specialist models.
A bigger future for the sisters.
SOLANA CONTRACT ADDRESS
View $HFAI on pump.fun 5RodQQJ5e3kv4z3E8P8E3wo79fSRZBRr43Ht455hpump2YEARS OF BUILDING
Four sisters. Two years of work.
A future we’re building every day.
We’ve spent two years bringing HeartFlowAI to life. The sisters, desktop workflows, and co-op experience are the foundation. $HFAI is our proposed funding engine for the next chapter: dedicated compute and models tuned around what each sister actually does.
OUR AMBITION Make $HFAI the greatest AI ecosystem token Solana has ever seen—and earn that place through products, useful skills, and releases people can inspect.
There is something to use.
Conversation, market analysis, execution, clarity, and a visible desktop companion. $HFAI grows from products people can explore and a team that keeps building.
Give every sister her specialty.
Direct realized $HFAI development funding into compute, curated data, model tuning, evaluation, and deployment. Measure improvements at each sister’s actual job.
Make participation useful.
Our proposed $HFAI direction connects community contributions, skill development, early feedback, and future governance to the next releases. Each feature needs its own published terms.
WHY $HFAI MATTERS NOW
The vision needs compute.
The compute cannot keep waiting.
We wanted to keep the token out of the spotlight a little longer. Compute is changing that timeline.
Two years of building have brought us here. Now we want dedicated hardware for repeated fine-tuning experiments, evaluations, and local inference. The longer we wait, the less predictable that hardware budget becomes. $HFAI is intended to help turn community momentum into the compute the sisters need next.
Our priority is to secure suitable 128GB systems as soon as funding and verified quotes allow, at the lowest total delivered cost we can find. We will compare NVIDIA and partner offers, check warranty and availability, and benchmark before scaling.
THE COST OF WAITING, SO FAR+74%
The 128GB Founders Edition’s US listed price is about 74% above its original MSRP. NVIDIA attributed the February adjustment to memory supply constraints.
This documents a rise that has already happened. Future prices could rise or fall; the 64GB announcement does not establish another 128GB increase.
More expensive. Same memory.
US list prices · checked| Recorded point | USD · bars start at $0 |
|---|---|
| Original MSRPBefore Feb 2026 | |
| First increaseFeb 2026 | |
| Current listing | |
Three documented price points, not a continuous price history or forecast. The current NVIDIA listing shows out of stock. Partner and regional prices, taxes, and delivery costs vary.
ANNOUNCED 2 OCT 2026A new 64GB Spark.
A new 64GB Spark.
A $4,999 starting price.
NVIDIA announced the 64GB partner model on October 2, with availability planned for October 23. Its starting price is higher than the 128GB model’s former $4,699 MSRP; the current 128GB listing is $6,950. That is the budget pressure behind our compute-first plan.
Read NVIDIA’s 64GB announcementFIRST PROPOSED COMPUTE MILESTONE
One 128GB development node.
A DGX Spark or compatible GB10 partner system for local inference, suitable fine-tuning workloads, and repeatable evaluations. Further nodes follow measured demand and budget.
Hardware planning reference · 1 unit$6,950 USD
Based on the checked NVIDIA listing, not a supplier quote. Final budget includes taxes, shipping, and any additional equipment.
FOLLOW THE WORK $HFAI FUNDS
Our proposed reporting cycle: publish the purchase budget, record actual hardware spending, share benchmarks for each sister, then show the model improvements in a versioned release. Procurement and funding progress will be reported when confirmed.
PUBLISHED $HFAI ALLOCATION FRAMEWORK
The treasury behind the vision.
Total supply in the published plan1,000,000,000$HFAI tokens
Project-owned allocation in the plan469,123,956Approximately 46.91% of total supply
Basis of the percentages belowProject allocation100% of the 469,123,956-token allocation
35%Development
The largest allocation backs the work that makes the sisters better.
| Purpose | Share | Tokens |
|---|---|---|
| Development | 35% | 164,193,385 |
| Marketing & partnerships | 15% | 70,368,593 |
| Reserve | 21% | 98,516,031 |
| Community rewards | 11% | 51,603,635 |
| Future use | 10% | 46,912,396 |
| Team | 8% | 37,529,916 |
Figures reproduce the published GitBook framework. Current on-chain balances, circulating supply, wallet controls, and vesting have not been verified here.
PROPOSED DEVELOPMENT PRIORITY
Train their strengths.
Keep their identities.
Start with capable open models, then fine-tune role-specific models or adapters on curated, licensed, and explicitly consented data. Evaluate each sister on her own tasks before shipping a versioned release.
AuroraCONNECTION & TOOL USETrain for natural dialogue, useful clarification, context handling, and dependable tool calls. Measure whether she understands the request and completes the intended task.
GigiMARKET ANALYSISTrain for source-grounded on-chain summaries, risk flags, and clear uncertainty. Measure factual accuracy and how well her explanations reflect the data.
NovaPLANNING & EXECUTIONTrain for task decomposition, prioritization, and practical follow-through. Measure plan quality, recovery, and completion on realistic tasks.
ClaraCLARITY & REGULATIONTrain for supportive dialogue, reflection, and consistent boundaries. Evaluate the usefulness, tone, and reliability of her responses.
01 Curate the data02 Tune the models03 Test on held-out tasks04 Release & improve
HOW $HFAI CAN FUND THE WORK
Turn realized funding
into better releases.
$HFAI treasury holdings must become realized funds before they can pay for hardware, data, or engineering. The proposed funding mix is disclosed treasury sales, project NFT proceeds where allocated to development, grants, and future product revenue. Buying $HFAI on a secondary market does not automatically send money to the development treasury.
Our proposed operating rules are public treasury addresses, multisig approvals, milestone budgets, and regular spending reports. Funding comes in, work ships, results are reviewed, and the next budget follows. These controls and revenue routes remain proposals until implemented.
CROSS-CHAIN DIRECTION · PROPOSED
Solana is the home.
The ecosystem can go further.
We plan to evaluate Wormhole Native Token Transfers to connect $HFAI across supported chains. A possible route is a Solana hub: lock tokens at the origin, issue an equivalent amount on a destination chain, and reverse that process when tokens return. This connects a shared $HFAI economy while accounting for supply across chains.
ORIGINSolana$HFAI locked
TRANSFER LAYERWormhole NTTVerified transfer messages
DESTINATIONSelected chainsEquivalent $HFAI issued
The route depends on token compatibility, mint permissions, selected networks, testing, rate limits, and a security review. No HeartFlowAI bridge deployment is claimed. WaifuVerse NFT transfers would need a separate implementation. Read Wormhole’s NTT documentation.
THE VISION IS BIG. THE WORK IS CONTINUOUS.The future of AI on Solana
The future of AI on Solana
should have a heartbeat.
Two years in, and we’re still building. More capable sisters. Better models. A stronger community. Every next step has to make HeartFlowAI more useful—and we intend to keep earning it.
Explore WaifuVerse11 / HFAI CLOUD · API
Your models.
Your wallet. One gateway.
CHAT · API KEYS · SOLANA PAY
Chat with the sisters in the browser.
Ship with discounted frontier-model APIs.
BUILT FOR THE SOLANA COMMUNITYChatBrowserAPI gatewayOpenAI-compatibleSolana PayPrepaid creditsFrontier modelsDiscounted
OPEN THE APP · WALLET SIGN-IN
HFAI Cloud
HFAI Cloud
in its own window.
A full application with a sidebar for Chat, Models, API keys, Usage, and Credits. Connect your Solana wallet once at the door — every section stays behind the login.
- Chat with all four sisters using your choice of frontier model.
- Create API keys for the OpenAI-compatible gateway.
- Top up prepaid credits with Solana Pay.
Early service · interactive preview until the API host is connected. See API.md.
12 / WHAT’S NEXT
Built with heart.
Improved with evidence.
PROPOSALS, NOT DATES
Working desktop navigation today.
Broader app and game coverage ahead.
00
DELIVERED
Current baselineDesktop navigation & co-op play
Mac, Linux, and Windows navigation, plus supported co-op integrations. Windows v12 supplies the documented test baseline.
Reliability
More app fixtures, measured task completion, recovery from focus changes, and clearer failures.
Voice
Evaluate local wake words, measure transcription and latency, explore speech interruption and usage controls.
Execution
Scoped permissions, stronger policies, file diffs, action previews, recovery, and independent injection testing.
Specialist intelligence
Curated role-specific training data, tuned models or adapters for each sister, and evaluations tied to real tasks.
Distribution
Signed releases, installers, updates, and more consistent distribution across Mac, Linux, and Windows.
Experience & ecosystem
Accessible transitions, voice-first onboarding, sister skill patches, Gigi signal upgrades, broader co-op game coverage, and Wormhole feasibility.
Longer research track: optional visual observation with explicit data handling, modular personalities, voice and VR, and embodied forms after the cognition is good.
13 / HEARTFLOWAI WAIFUVERSE
Meet the world
behind the sisters.
WaifuVerse brings the HeartFlowAI character universe into Solana collectibles. Explore the collection, its artwork, and the community growing around it.
Agent-linked identity and access are proposed features. Availability and eligibility will be defined in future releases.
THE NEXT CHAPTER IS A CONVERSATION.
Come meet your sisters.
Follow the work. Try a conversation. Help shape what comes next.