Ochob

Dev making Web3 moves | Crypto nerd & builder on @base
Betting on @polymarket

[Ochob @0chob]

https://t.co/IMNzYFJXZ6

[Ochob @0chob]

@misat0x Don't tell them. They're happier this way

[Ochob @0chob]

Four aluminum boxes on a desk. Same power circuit as his coffee machine. Running a trillion-parameter model in silence.

Four Mac Studios. 512GB of unified memory each. Linked into one 2TB machine. Kimi K2 Thinking - a frontier open-weight model - running locally, fast enough that the demo doesn't need a loading spinner.

The rig cost less than one year of what a mid-size agency spends on OpenAI. No cloud bill. No rate limits. No client data leaving the room.

While the competition argues about API tiers, he's selling outcomes. Contract review that finishes overnight. Onboarding flows that used to need three humans. Lead scoring built on the client's actual pipeline data instead of a generic model's guess.

$5,200 a month per client. No usage caps because there's no usage meter. Savings cover the invoice inside three weeks. Nobody has churned.

7 clients paid for the desk. The desk is why the next 30 don't scare him.

[Orion @0xOrionVega]

https://t.co/vKscXpdeMQ

[Ochob @0chob]

@velesxbt wait, I thought it could do stuff like this. What am I actually paying $4k for then?

[Ochob @0chob]

@0xLand0 right sized, everything else is cosplay

[Ochob @0chob]

Do you actually understand what's happening here. He built a full virtualization stack - Proxmox host, Windows jump box, Pi, 6-node ARM cluster, router, 10-gig switch - into a rack the size of a shoebox. And now he's arguing with himself about 1/3 of a rack unit.

The current build lives in the DeskPi Rackmate T1, an 8U mini rack that fits on a shelf. Inside: 1 Proxmox virtualization host, 1 Windows jump box on a mini PC, 1 Raspberry Pi, a 6× CM4 cluster board (6 ARM nodes on one PCB), a router, and an 8-port switch with 2 SFP+ cages. That's roughly 11 machines in a footprint smaller than a microwave.

The next project is the Rackmate T0. Four rack units of usable space. A "micro rack" that makes the T1 look enterprise. He's cleaning up the T1's cable spaghetti by installing a conference-table pop-up power unit as the main PDU. 3D-printed rack ears. Chassis notched to accept it.

One catch. The unit measures 1.33U. Rack rails come in 1U increments. Do you eat a full 1.33U at the bottom and leave a rear strip of empty space? Or top-mount on one row of holes and let the bottom float and flex? He's asking.

The people writing homelab guides on YouTube are still explaining what a rack unit is. He's already inventing hardware categories to solve problems that don't have names yet.

8U. 4U. 1.33U. Mini rack. Micro rack. What comes next - nano rack?

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

He stopped renting AI. $147 in eBay parts stacked in a shoebox, one weekend of work, and now three agents work for him while he sleeps for $6 a month in electricity.

The box used to hold a pair of sneakers. Now it holds a used RTX 3060, a cheap mini-ITX board, and three local models running 24/7 on his own hardware, his own weights, his own terms.

Two commands to boot. Zero recurring fees. Zero API keys.

Gatekeeper reads his inbox overnight, sorts by urgency, drafts five replies. He wakes up, approves four, sends them from bed before coffee.

Cartographer scrapes every article, thread, and PDF he saves during the day, pulls out what matters, and files it into an archive he can query in plain English.

Watcher tracks one competitor's pricing page, one keyword on X, one job board. Silent until something moves. Pings his phone only when it's worth interrupting him.

Three jobs. Running while he sleeps. $6 a month, all in.

He used to pay $389 a month for the same output on someone else's servers. Now he pays $147 once, $6 a month forever, and the person deciding when the model changes is him.

The shoebox is the least interesting part of this story.

[Gipp 🦅 @gippp69]

https://t.co/oMgggyvOz6

[Ochob @0chob]

@0xDezo that's really smart move to make own workstation

[Ochob @0chob]

AWS Bedrock Lost A $340K/Year Healthtech Contract To Four Framework Desktops On A Wire Shelf In Hyderabad

vikram is 29, second floor of his parents' house, four framework desktops with ryzen ai max+ 395 on a wire shelf next to the water tank

pause at 0:22 on the shelf shot, four identical 128gb unified memory boxes wired into one 10gbe switch and a shoebox ups, that is the entire $340k/year infrastructure

$28,400/month on the anchor contract, two shorter engagements at $9,000 and $14,000, $47 in telangana power, hardware paid for in 11 days, qlora runs on llama 3.3 70b overnight and by morning the client has a domain-tuned checkpoint that never touched a us cloud

BAA covers inference, not the training loop, the aws shops spent four months trying to word around that and couldn't

the moat is not price, it is a clause his architecture answers and theirs can't

follow before their compliance teams find him

[Ochob @0chob]

AWS Bedrock Lost A $340K/Year Healthtech Contract To Four Framework Desktops On A Wire Shelf In Hyderabad

vikram is 29, second floor of his parents' house, four framework desktops with ryzen ai max+ 395 on a wire shelf next to the water tank

pause at 0:22 on the shelf shot, four identical 128gb unified memory boxes wired into one 10gbe switch and a shoebox ups, that is the entire $340k/year infrastructure

$28,400/month on the anchor contract, two shorter engagements at $9,000 and $14,000, $47 in telangana power, hardware paid for in 11 days, qlora runs on llama 3.3 70b overnight and by morning the client has a domain-tuned checkpoint that never touched a us cloud

BAA covers inference, not the training loop, the aws shops spent four months trying to word around that and couldn't

the moat is not price, it is a clause his architecture answers and theirs can't

follow before their compliance teams find him

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

A father in Aarhus spent three weekends making sure a corporation would never be the voice in his daughter's bedroom again.

So he became it himself.

For eight months, his 8-year-old has been talking every night to a teal cube on her nightstand. It looks like BMO from Adventure Time. It answers in 1.8 seconds. It has 3 billion parameters. It has never once been connected to the internet - and the voice coming out of it is her dad's, pitched up and filtered to sound like the show. She knows. That's the whole point.

Her father, Mads Holberg, 3D-printed the BMO shell in two pieces, then put a Raspberry Pi 5, a four-mic array, and a speaker inside it. She asks BMO why the sky is blue. BMO answers in his voice. Amazon never hears the question. Google never sees the answer. And the last thing she hears before she falls asleep belongs to her father, not a server farm in Virginia.

Pause at 0:11.

Listen to the cadence. It's pitched up and processed to sound 8-bit, but the rhythm underneath is wrong for a cartoon - it's the rhythm of a man reading to a child. That's because it is. Mads recorded 90 minutes of himself reading her favorite books, trained a Piper voice on the audio, and ran it through a filter chain that makes him sound like BMO. Every answer she hears is, in some small way, him.

The shell is printed off Printables. Inside, a Raspberry Pi 5 with the Wi-Fi module physically removed - the device cannot reach the internet even if it wanted to. Whisper handles listening, a quantized open model handles thinking, his own voice handles the talking. A four-mic array hides behind the side grille. The right eye is a camera, so she can hold up a leaf and ask what kind of tree it came from.

Total parts cost: €184. Total recurring cost: zero.

Then a parent from her class came over for a playdate, watched the cube for ten minutes, and asked if Mads could build one for her son.

He's now made 23 of them.

Each one comes empty. The parent records their own voice - 90 minutes of whatever bedtime book they want - and Mads trains the model on it before shipping. He charges €420. There's a waitlist of 60. Three of the buyers are grandparents recording for grandchildren they don't see often enough. One is a mother with stage 3 cancer.

His wife asked him why he didn't just buy an Echo Dot eight months ago.

He opened the Amazon privacy dashboard for the Echo they used to own and showed her the voice clip archive - over 14,000 recordings of their daughter, going back to when she was six. He unplugged that device the week she turned seven.

Then he spent three weekends making sure the next voice in his daughter's life would be his.

Now he's doing it for other parents too.

That's the gap. Big Tech sells you a microphone in a speaker. He's selling parents a way to be the voice in their children's rooms - even when they can't be.

[Ochob @0chob]

A pediatric clinic chain in Munich pays Tomasz Kowalski 9,200 euros a month for an AI rig he builds in his garage in Olsztyn and ships across the border.

Pause at 0:11. He's installing the motherboard but skips one step every PC builder does. He doesn't attach the Wi-Fi antennas. The M.2 slot for the wireless card stays empty. The machine he's building physically cannot join a wireless network.

Their compliance officer wouldn't sign off on cloud AI. German medical data law made every major vendor impossible. ChatGPT, no. Claude, no. Gemini, especially not.

So Tomasz pitched the one thing nobody else did: a workstation that physically can't phone home.

Intel Core Ultra 7 265K. RTX 5070. 96GB DDR5. 4TB NVMe with no cloud sync. No Wi-Fi card. No Bluetooth module. One ethernet port, locked to the clinic's internal VLAN.

Six agents run on it inside the clinic's server closet. Insurance pre-auth calls. Parent reminders in DE, TR, AR. Symptom triage on intake forms. ICD-11 billing code suggestions. Vendor invoice reconciliation. Daily summaries for the doctors.

The clinic used to pay a Munich BPO 31,000 euros a month for the first four. Now they pay Tomasz 9,200.

But the savings aren't the moat. The moat is that the box can't reach the internet without someone physically plugging in a cable the clinic's IT controls. Patient data never touches a US server. Never trains anyone else's model.

Every operator is selling speed and cost. Tomasz sold sovereignty.

The contract is signed for five years.

Most people are racing to the bottom on price. The smart ones are racing to the air gap.

[Ochob @0chob]

A father in Aarhus spent three weekends making sure a corporation would never be the voice in his daughter's bedroom again.

So he became it himself.

For eight months, his 8-year-old has been talking every night to a teal cube on her nightstand. It looks like BMO from Adventure Time. It answers in 1.8 seconds. It has 3 billion parameters. It has never once been connected to the internet - and the voice coming out of it is her dad's, pitched up and filtered to sound like the show. She knows. That's the whole point.

Her father, Mads Holberg, 3D-printed the BMO shell in two pieces, then put a Raspberry Pi 5, a four-mic array, and a speaker inside it. She asks BMO why the sky is blue. BMO answers in his voice. Amazon never hears the question. Google never sees the answer. And the last thing she hears before she falls asleep belongs to her father, not a server farm in Virginia.

Pause at 0:11.

Listen to the cadence. It's pitched up and processed to sound 8-bit, but the rhythm underneath is wrong for a cartoon - it's the rhythm of a man reading to a child. That's because it is. Mads recorded 90 minutes of himself reading her favorite books, trained a Piper voice on the audio, and ran it through a filter chain that makes him sound like BMO. Every answer she hears is, in some small way, him.

The shell is printed off Printables. Inside, a Raspberry Pi 5 with the Wi-Fi module physically removed - the device cannot reach the internet even if it wanted to. Whisper handles listening, a quantized open model handles thinking, his own voice handles the talking. A four-mic array hides behind the side grille. The right eye is a camera, so she can hold up a leaf and ask what kind of tree it came from.

Total parts cost: €184. Total recurring cost: zero.

Then a parent from her class came over for a playdate, watched the cube for ten minutes, and asked if Mads could build one for her son.

He's now made 23 of them.

Each one comes empty. The parent records their own voice - 90 minutes of whatever bedtime book they want - and Mads trains the model on it before shipping. He charges €420. There's a waitlist of 60. Three of the buyers are grandparents recording for grandchildren they don't see often enough. One is a mother with stage 3 cancer.

His wife asked him why he didn't just buy an Echo Dot eight months ago.

He opened the Amazon privacy dashboard for the Echo they used to own and showed her the voice clip archive - over 14,000 recordings of their daughter, going back to when she was six. He unplugged that device the week she turned seven.

Then he spent three weekends making sure the next voice in his daughter's life would be his.

Now he's doing it for other parents too.

That's the gap. Big Tech sells you a microphone in a speaker. He's selling parents a way to be the voice in their children's rooms - even when they can't be.

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

A pediatric clinic chain in Munich pays Tomasz Kowalski 9,200 euros a month for an AI rig he builds in his garage in Olsztyn and ships across the border.

Pause at 0:11. He's installing the motherboard but skips one step every PC builder does. He doesn't attach the Wi-Fi antennas. The M.2 slot for the wireless card stays empty. The machine he's building physically cannot join a wireless network.

Their compliance officer wouldn't sign off on cloud AI. German medical data law made every major vendor impossible. ChatGPT, no. Claude, no. Gemini, especially not.

So Tomasz pitched the one thing nobody else did: a workstation that physically can't phone home.

Intel Core Ultra 7 265K. RTX 5070. 96GB DDR5. 4TB NVMe with no cloud sync. No Wi-Fi card. No Bluetooth module. One ethernet port, locked to the clinic's internal VLAN.

Six agents run on it inside the clinic's server closet. Insurance pre-auth calls. Parent reminders in DE, TR, AR. Symptom triage on intake forms. ICD-11 billing code suggestions. Vendor invoice reconciliation. Daily summaries for the doctors.

The clinic used to pay a Munich BPO 31,000 euros a month for the first four. Now they pay Tomasz 9,200.

But the savings aren't the moat. The moat is that the box can't reach the internet without someone physically plugging in a cable the clinic's IT controls. Patient data never touches a US server. Never trains anyone else's model.

Every operator is selling speed and cost. Tomasz sold sovereignty.

The contract is signed for five years.

Most people are racing to the bottom on price. The smart ones are racing to the air gap.

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

He doesn't sell software. He sells a black box.

$1,480 in parts. Intel Core Ultra 7 265K, RTX 5070, 64GB DDR5, a cheap UPS. The box lives on a shelf in the client's server closet. The client is never allowed to open it.

Pause at the second photo - one Tailscale sticker, no other markings.

Marek is 27. Walks into small-to-mid businesses around Indianapolis and asks one question: "what does your team do, every single day, that's basically copy-paste?"

He writes down the answer. Goes home. Spends a weekend wiring six local agents on the box - email triage, invoice OCR, lead enrichment, support drafts, vendor follow-ups, meeting notes.

Installs it on a Monday. Hands the client a Tailscale link and a $2,500/month invoice.

One client was paying $28,400/month for three coordinators doing exactly that work. They're now paying Marek $2,500. Net savings: $25,900/month. The CFO did the math once and stopped asking questions.

Eight clients on the same setup. $20,000/month MRR. ~85% margin. Only fixed costs are electricity and one offsite encrypted backup.

Here's what makes it unkillable: the data never leaves their building. No OpenAI calls. No Anthropic logs. The law firm client, the accountant, the medical billing shop - none of them are legally clean sending half this work to a cloud model. Marek isn't competing with ChatGPT. He's competing with the cost of hiring another human.

They will not cancel. The math won't let them. Compliance won't let them either.

Everyone in his city is still arguing about whether the next model will be AGI. Marek is closing client number nine.

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

@Di_Krass_ It’s actually a multi-agent AI system. Each one has its own brain, personality, and tasks

[Ochob @0chob]

That's 9 Stack-chans wired to Claude Code - each toy desk-pet is now the face of a parallel agent running on the same MacBook.

Pause at 0:17.

Stack-chan is an open-source ESP32 robot. Cute swinging head, pixel face, normally used as a $60 desk companion that says good morning. Operator in Osaka bought 9 of them, flashed custom firmware, and assigned one per Claude Code agent.

The faces aren't decoration - they're the status UI. Each Stack-chan changes expression based on what its agent is doing:

flat mouth = idle, ready for a task
squinting eyes = thinking
straight line = blocked, needs input
closed eyes + smile = task done
The red device in his hand is a physical dispatcher. He scans the row, finds a Stack-chan that's smiling (idle), clicks once, and the task drops into that agent. No tabs, no terminal-switching, no "which window was the auth fix in."

Why he built it: he was running 6-9 parallel Claude Code agents and losing 2-3 hours a day just figuring out who was stuck. So instead of reading logs, he gave every agent a face.

Last week:

43 PRs merged, up from 11 the week before
~2.5 hours/day saved on status-checking
0 silent failures - the sad face catches what he used to find at 11pm
Total cost of the rig: ~$540 in Stack-chans + a $42 MDF plank.

He's not coding faster. He's just not paying the 30-second context-switch tax every time he goes hunting across 9 terminals.

The whole thing works because Stack-chan was already built to have a face. He didn't invent the hardware - he just realized a toy robot is a better status indicator than a Slack notification.

That's the gap. Everyone running parallel agents drowns in tabs and tells themselves it's fine. He bought 9 desk toys, flashed them, and now runs 9 agents the way most people run one.

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

A guy in Lagos pulls $11,840 a month from a box that fits in his palm.

No GPU. No cloud bill. No server rack humming in a closet somewhere.

Just a Raspberry Pi 5, a Samsung 990 PRO NVMe, two whisper-quiet fans, and a Pironman MAX case he assembled himself one weekend. Full setup cost him under $400.

The build is doing one thing all day: running 23 Claude Code agents in parallel, each one wired to a different client's repo through MCP.

Pause at 0:08.

That orange ribbon cable snaking out the top? That's not display. That's a second board sandwiched underneath, running the orchestrator that fans incoming jobs out to the right agent. He flashed the OS and routed the cooling himself in an afternoon.

Each agent has a narrow job. One refactors legacy PHP for a dental SaaS in Tampa. One writes Shopify Liquid for a print-on-demand store in Manchester. One drafts weekly changelogs for a fintech in Singapore. None of them know about each other. None of them need to.

He charges between $290 and $740 per repo per month. Right now he has 28 active.

The interesting part isn't the money. It's the form factor.

For two decades the rule was: more compute, more colocation, more rent. The whole industry was built around the idea that serious work required serious infrastructure.

This guy serves 28 clients from a cube that draws less power than a desk lamp.

When your office is your palm, the old questions stop making sense. No team to manage. No overhead to cover. No rent to chase. You just need the box to keep running and the agents to keep shipping.

The future of work isn't remote.

It's about the size of a sandwich.

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

A 19-year-old is making $164,000/month from creators who speak Spanish, Portuguese, Italian, Korean, and English.

She only speaks one of them.

Pause at 0:22.

The screen recording opens on what looks like a college girl's bedroom. A psychology textbook face-down on the desk. A half-finished iced coffee. A Spotify playlist labeled "study" playing in the corner. The monitor on the desk has 5 tabs open across the top.

Each tab is an Instagram account.

Tab 1: a Spanish-language clean-girl beauty creator based in Mexico City. 412,000 followers. Her latest reel is a nighttime skincare routine. The caption is in Spanish.

Tab 2: an Italian-language minimalist fashion creator based in Milan. 287,000 followers. Latest reel is an outfit-of-the-day. The caption is in Italian.

Tab 3: a Portuguese-language fitness creator based in Rio. 519,000 followers. Latest reel is a beach workout. The caption is in Portuguese.

Tab 4: a Korean-language K-beauty creator based in Seoul. 334,000 followers. Latest reel is a 10-step skincare routine. The caption is in Hangul.

Tab 5: an English-language "soft girl" aesthetic creator based in Brooklyn. 198,000 followers. Latest reel is a morning routine. The caption is in English.

She's in her dorm in Boston. She's a sophomore. She speaks English and exactly 14 words of high school Spanish.

The hands on every reel are hers.

Same nail shape. Same ring on the right middle finger. Same gesture at the 11-second mark of each video — a small twist of the wrist when she introduces the product. Five different bedrooms. Five different lighting setups. Five different aesthetic systems. One pair of hands across all of them.

She doesn't write the captions. Claude does. Each account has a 3,200-word system prompt that locks the voice in its native language - the Mexican creator uses regional slang, the Milan creator uses the restrained vocabulary of Italian luxury fashion, the Rio creator uses Carioca Portuguese, the Seoul creator uses Korean honorifics correctly, the Brooklyn creator uses the chill American voice.

Claude writes 30-40 captions a week across all 5 accounts. ElevenLabs handles the voiceovers in each language - voice clones trained on 60-second audio samples she paid native speakers to record on Fiverr for $40 each.

She built the first persona 11 months ago. The Brooklyn one. Standard play - Higgsfield Soul ID, LoRA training, 90 source images, the usual stack. By month 3 the Brooklyn account was hitting 80K followers and her first brand deal closed at $1,800.

Then she did something almost nobody in this market does.

She didn't build a second persona in English. She built one in Spanish.

The insight was this: the AI UGC market in the US is saturated. Mid-tier brands get pitched by 60+ creators a week, and the conversion rate on cold outbound is dropping every quarter. Spanish-language brands in Mexico and Latin America have almost no AI creator presence - the market is wide open. Italian and Portuguese are even more underserved. Korean has the highest per-deal pricing in the world for skincare UGC because the K-beauty industry is massive and willing to pay.

She speaks none of these languages. She doesn't have to.

Claude handles the writing. ElevenLabs handles the voice. DeepL handles the contracts. A Cambly tutor she pays $24/hour handles cultural review - she'll send finished reels to a native speaker tutor and ask "does this read as natural in your country?" before posting. Roughly 4 hours/week of tutor time across all 4 non-English languages.

When a brand DMs the Milan creator and the conversation needs to go beyond Claude's autoresponder, she routes it to a separate Cambly tutor — she has 5 of them on retainer, one per language — who closes the deal as the persona's "manager." The tutors think they're freelance translators helping a young entrepreneur manage international clients. They don't know the creators aren't real.

Active brand contracts across the 5 personas:

Mexico creator: rolling retainer with a Mexico City skincare brand, $4,200/month
Milan creator: 6-month exclusive with an Italian boutique fashion house, €38,000 over term
Rio creator: 12-month with a Brazilian activewear brand, R$240,000 over term (~$48K USD)
Seoul creator: 3-month K-beauty campaign with a Seoul-based sheet mask brand, ₩52M (~$38K USD)
Brooklyn creator: rolling retainer with a US skincare DTC, $6,800/month

Total in-flight contract value across the 5 personas, converted to USD: roughly $387,000 over the next 12 months.

Last month's invoiced revenue: $164,000. The month before: $138,000. The month before that: $109,000.

Her monthly stack cost: $720 - Higgsfield Pro across 5 character licenses, Claude API at heavy usage, ElevenLabs Pro for 5 voice clones, DeepL Pro, Cambly tutor retainers, and the Lovable dashboard that orchestrates the whole thing. Margin: 99.6%.

Someone in the replies asked the obvious question. Why doesn't she just hire 5 real creators in 5 countries? It would be simpler.

Her response in a quote tweet: "I'd be running a talent agency. Talent agencies have margins of 12%. I have margins of 99%. The creators aren't the bottleneck - they're the constraint. Removing them is the whole edge."

The geographic arbitrage compounds. Brand deals in Italy and Korea pay 15-30% more per follower than US deals because the local AI creator supply is so thin. Her Milan creator at 287K followers commands per-post rates that a US creator at the same size couldn't get. Her Seoul creator at 334K is being offered K-beauty brand exclusives that would normally go to creators with 800K+.

Someone tried to replicate the model. They built 3 personas in 3 languages, hired 3 freelance translators, pitched 40 brands. Closed two deals at small budgets. They posted a teardown asking what they were missing.

She replied: "You hired translators. They're not the same as cultural reviewers. A translator gives you accurate language. A cultural reviewer tells you when your Italian fashion creator is using Spanish flatness or your Korean creator is being too direct for the culture. The product isn't the language - it's the cultural fit. That's why the captions take me an hour each and not five minutes."

She's 19. She's been doing this since November of last year. She has midterms next week. Her current monthly run-rate puts her on track to clear $2.4M by her 20th birthday.

She's never been to Mexico, Italy, Brazil, or South Korea.

Somewhere, a 28-year-old human creator in Milan with 90,000 followers is wondering why brands keep ghosting her this quarter. Why every campaign she pitches says "we've already locked our influencer for that market."

She isn't competing with bigger creators.

She isn't competing with local talent agencies.

She's competing with a sophomore in Boston who speaks 14 words of Spanish.

[Ochob @0chob]

@0xOrionVega Right? Behavior change, not workflow change. You stop hovering

[Ochob @0chob]

$36 plastic toy. $28,000/month consulting income.

A dev at the Anthropic conference grabbed the free Cardputer keychain, walked home, and turned it into the device he now runs his entire Claude Code business from.

Look at what he built into something the size of a credit card.

The hardware:

1 M5Stack Cardputer, retail $36, free at the conference
1 thumb-sized QWERTY keyboard
1 1.14" color screen
1 built-in mic and speaker
1 WiFi chip that talks to his Mac across the apartment

He hasn't opened his laptop screen in 11 days.

The four modes he wired up:

Remote terminal. He types a question on the Cardputer's tiny keyboard. WiFi pushes it to Claude Code running on his Mac in the other room. Claude executes. The result streams back to the 1.14" screen. He reads it from the couch.

Voice intercom. Hold the side button, talk into the mic, release. The clip uploads to his Mac, runs through a local STT model, gets fed to Claude. The answer comes back as audio out of the Cardputer's speaker. Press to talk, release to hear. Like a walkie-talkie to an AI.

Physical pet mode. A pixel-art animal lives on the screen. It walks around when Claude is thinking. It looks up and speaks when Claude is answering. He always knows what state the agent is in without checking his laptop. A tamagotchi for agent status.

OpenClaw passthrough. Plug the Cardputer into the OpenClaw rig and it becomes the voice input layer for everything. One device, every Claude entry point.

Everything happens locally except the actual Claude call. No cloud STT. No cloud TTS. The Cardputer is dumb hardware. His Mac does the work. WiFi carries the packets.

Client work used to mean sitting at his desk for 11 hour blocks. Now he ships from his couch, the kitchen, the balcony. Same deliverables. Same rate. 4x the throughput because he can think while Claude works and check status from anywhere.

He says the pet mode is the one that changed his business. Not the terminal. Not the voice.

Because the second you can see Claude thinking from across the room, you stop hovering and start billing for the next thing.

Most devs are chained to the same screen as the agent, billing one project at a time. He's on the couch with a $36 keychain, running three in parallel.

That's the gap.

[Ochob @0chob]

https://t.co/22XbJRZsvG

[Ochob @0chob]

@sairahul1 It's really not that hard with the AI capabilities