Your Secrets Stay Home: Underdog, the Assistant That Never Leaves Your Laptop

The problem

Think about what you would need to hand over for an AI assistant to really run your life. Your inbox. Your calendar. Your meeting notes. Your bank and shopping accounts. The more it knows, the more useful it gets.

Now think about where all of that goes. With almost every assistant today, it leaves your computer and lands on a company’s servers. You are trusting that company, its staff, its security and its future business plans with the most private parts of your life.

The founder of Underdog learned this the hard way. A bug in a popular email app exposed his private emails to other users. The data was “safe in the cloud” until it wasn’t.

So the real question is: can you have a capable AI assistant without shipping your private life to someone else’s computer?

What we’re solving

Underdog is a personal AI assistant from Conway Research, a new AI lab founded by Sigil Wen. It opened as an invite-only beta on 5 October 2026. The whole pitch fits in one sentence: the AI runs on your own computer, so your data never has to leave it.

What it takes off your plate

• Your data stays home: The AI model runs entirely on your device. No server reads your emails, transcripts or calendar to answer you.
• No subscription, no ads: Underdog is free at launch, and Wen says it will never show ads.
• It works offline: Investor a16z describes it as helping you even with no internet connection.
• Locked account keys: The keys Underdog uses to reach your email and other accounts are encrypted.

What it does today

• Email (Gmail and Outlook), calendar, meeting transcription and dictation, all on the device.
• Everyday tasks like shopping research and homework questions.
• A preview tool for asking data brokers to remove your details.

What it does not solve

• The services you connect still see their part. Running the AI locally takes the AI company out of the picture. It does not stop Gmail, a shop or a bank from seeing what passes through them.
• It is not on your phone yet. Mac and Windows only for now. Linux, iPhone and Android are promised with no dates.
• It is not the smartest AI out there. Its models are much smaller than the giant cloud models.

A real example

Underdog has been in a closed beta for about a week, so there are no independent user stories yet. Its launch listing on Product Hunt had a single upvote and one comment when I checked. Instead of stretching a thin example, here is how Underdog compares with the other personal agents people are actually choosing between.

Who is behind it

• Founder: Sigil Wen, a self-taught coder and Thiel Fellow who moved to Silicon Valley at 17. He once got GPT-2, an early OpenAI model, running on an Apple Watch.
• Track record: He was among the earliest testers of tools that became Claude, Midjourney and Stable Diffusion, and Naval Ravikant hired him to work on Airchat.
• Backers: Andreessen Horowitz (a16z) led the first round. Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund joined, plus Stripe co-founder Patrick Collison as an angel. The amount was not disclosed.
• His bet, “Underdog’s Law”: Wen argues that today’s top AI reaches ordinary devices within about six months. It is a forecast, not a measured fact.

How it works (with code)

Think of the difference between a restaurant and your own kitchen. Cloud AI is the restaurant: the food is great, but you hand your order to strangers and eat on their premises. Underdog is a good home kitchen: a smaller menu, but nothing leaves your house.

The pieces

• Two brains: Woof is a small, fast model with 4 billion parameters that takes under 2.5 GB. Underdog 27B is a bigger model with 27 billion parameters for harder tasks, built on Alibaba’s open Qwen3.8-27B model.
• Parameters, in plain words: The numbers a model learned during training. More parameters usually means smarter, but also more memory.
• Husky, the engine: Wen built his own software to run models quickly on your hardware. He says it moves less data between your computer’s main chip and its graphics chip than rival engines.
• A locked vault per account: Each connected account gets its own encrypted vault, with the key kept on your computer, according to its launch listing.
• A “Naval mode” switch: The listing says it can stop all outbound calls home from the app.

Will it run on my laptop?

Running AI at home is mostly a memory question. A model’s numbers have to fit in your computer’s memory (RAM). This short script estimates whether each Underdog model fits on a laptop with 8, 16 or 32 GB.

# Will a local AI model fit on my laptop? A rough estimate.
 
def weights_gb(params_billion, bits):
    # Each parameter is a number; 4-bit means half a byte per number
    return params_billion * bits / 8
 
def check(model, params_billion, bits, ram_gb, buffer_gb=4):
    # buffer_gb: rough room for the chat itself and your other apps
    need = weights_gb(params_billion, bits) + buffer_gb
    verdict = "fits" if need <= ram_gb else "too big"
    label = f"{model:<13}{bits}-bit" print(f"{label} ~{need:4.1f} GB | {ram_gb:>2} GB: {verdict}")
 
models = [("Woof 4B", 4, 4), ("Underdog 27B", 27, 4), ("Underdog 27B", 27, 8)]
for ram in (8, 16, 32):
    for name, size, bits in models:
        check(name, size, bits, ram)
    print()

Output when I ran it:

Woof 4B      4-bit ~ 6.0 GB |  8 GB: fits
Underdog 27B 4-bit ~17.5 GB |  8 GB: too big
Underdog 27B 8-bit ~31.0 GB |  8 GB: too big
 
Woof 4B      4-bit ~ 6.0 GB | 16 GB: fits
Underdog 27B 4-bit ~17.5 GB | 16 GB: too big
Underdog 27B 8-bit ~31.0 GB | 16 GB: too big
 
Woof 4B      4-bit ~ 6.0 GB | 32 GB: fits
Underdog 27B 4-bit ~17.5 GB | 32 GB: fits
Underdog 27B 8-bit ~31.0 GB | 32 GB: fits

• A model’s memory need is roughly its size times how many bits each number uses. “4-bit” is a compressed version that uses half a byte per number.
• The small Woof model fits on almost any recent laptop.
• The bigger 27B model needs a well-equipped machine. On a typical 16 GB laptop it does not fit, by this estimate.
• The 4 GB buffer is my rough allowance for the conversation and your other apps. Conway has not published official memory requirements, and which compression it uses for the 27B model is not public, so treat these as estimates.

How fast is it?

• Conway’s own tests show Husky running Woof at up to 730 tokens (small pieces of words) per second, and up to 4.5x faster than Apple’s own engine on one editing task.
• The catch: those numbers come from the small model on a top-end Apple M5 Max laptop with 128 GB of memory, tested by the company itself. Results on Windows may differ.

What it costs

Prices below were checked on 12 October 2026. Ollama’s prices come from its own pricing page. Underdog’s terms come from its founder’s statements to TechCrunch. Meta Muse and typical chatbot prices are third-party reports. All amounts in US dollars.

• Free is not free. With Underdog, you pay with hardware and battery instead of a subscription. Running AI at home uses your computer’s memory, power and fans.
• Watch the business model. Conway earns only when the assistant spends money for you. The fee percentage has not been published.
• The hardware bill is real. The big 27B model wants roughly 32 GB of memory by my estimate. If your laptop has 8 or 16 GB, you get the small model only.
• Solo user: Try Underdog if you have an invite and a recent laptop. Keep a cloud chatbot for hard questions.
• Small team or business: Wait. It is a one-week-old beta with no published security audit.

Pros and cons

Pros

• Your private data stays on your device. No AI company reads your emails or notes to answer you.
• No subscription and no ads. Wen says his costs are low because your computer does the work.
• Works offline. Useful on flights or poor connections.
• Privacy extras. Encrypted account vaults, a switch to block outbound calls, and stripping of tracking pixels from email, per its launch listing.
• Serious backers. a16z, Khosla Ventures and Stripe’s co-founder are betting on it.

Cons

• Smaller brain than the cloud. Its models are far smaller than state-of-the-art cloud models. Wen’s claim that the 27B model matches Claude Opus 4.6 applies only to some unnamed tests, with no independent check yet.
• Speed claims need context. The headline speeds come from the small model on a top-end, 128 GB Apple laptop, measured by the company.
• Your battery pays. A vendor of rival AI chips estimates a large local model can drain a laptop battery in about an hour. That source has its own product to sell, but the direction is right: local AI is power-hungry.
• An awkward incentive. A company that earns only when you buy things has a reason to nudge you toward buying. That is the same kind of incentive problem it criticises in ad-funded apps.
• Local is not fully private. Gmail, shops and banks still see what you do with them. Only the AI middleman is removed.
• Early and unaudited. Invite-only, Windows still in beta, no phone app, and no external security audit reported.

Key takeaways

• Underdog is a free AI assistant that runs on your own Mac or Windows PC, so your emails, notes and calendar never go to an AI company’s servers.
• It makes money from a small, undisclosed cut of purchases it makes for you, not from ads or subscriptions.
• The trade-off is brains and hardware: its models are smaller than cloud AI, and the bigger one needs a well-equipped computer.
• Its speed and quality claims come from the company itself and have not been independently tested.
• If privacy matters most to you and you have a strong laptop, it is worth trying. If you want the smartest answers, cloud AI still wins today.