VibeAI OS How you payDocsPricing We run it for you
VibeAI OS // no subscription. no data harvesting.

Your Personal AI. Owned.

It knows your projects, answers on Telegram at 3 AM, ships real apps to real URLs — and earns money with your idle hardware. Yours, on your machine. Chat · Projects · Agents · Model Stack · Swarm

You pay the AI company directly. We add nothing. First call: 15 minutes, no pitch.

SWARM ONLINE one install · live in minutes 10 AI brands · your key, zero markup your data stays on your box built in-house, top to bottom
Agents that work for you

Your agents,
working for you.

your goalplansyou approvebuildsverifiesdone
building · idea → live URL00:00
"build me a booking site"
earning · crypto in, no tradingwallet · self-custody
0.00EURC
Inside the dashboard

Chat with it.
Or open the IDE.

The two surfaces you'll live in — exactly as they look on your dashboard. Click the tabs.

Advanced chatProject IDE
myshop/
  index.html
  pricing.tsx
  checkout.tsx
  styles.css
  tasks.md · 3 open
pricing.tsxagent working…preview: myshop.app… ● live

^ THE ACTUAL SURFACES — CHAT CONSOLE AND IDE, AS THEY LOOK ON YOUR DASHBOARD

Real pixels

No mockups.
Screenshots.

Captured from a running install of v0.8.13 — this week's release, untouched.

VibeAI OS dashboard — live system globe, sessions, services
#dashboard — the home screenv0.8.13 · live capture
Chat — the assistant plans your day, drafts an invoice, sets a reminder
#chat — your day, plannedreal UI · illustrative conversation
Chat, advanced use — tool events live: test run fails, file read, patch write-verified, 214 tests pass
#chat — find it & fix it · tool eventsreal UI · illustrative conversation
Models page — active model, thinking mode, request-kind routing
#models — pick & route brainslive capture
Wallet page — on-chain wallet, testnet pilot, assets
#wallet — testnet pilot, labeledlive capture
One OS · three jobs
YOUR PERSONAL ASSISTANTIt knows you. Locally.

"Brief me at 7." "What did we decide last week?" It remembers your preferences, projects, and corrections forever — and only your machine knows any of it.

Chat · Sessions · Channels · Cron · Voice · Memory

YOUR ENGINEERIt ships.

"Build me a booking site." Twenty minutes later there's a URL. Point at a button, say what to change, watch it redeploy — write-verified, Claude-Code-caliber. You review. It does.

Vibe Code · Projects · IDE · Autopilot · Research · Model Stack

YOUR EARNING MACHINEIt pays.

The wild part: when you're not using it, other machines hire it. Verified job in, money out — straight to a wallet only you hold. Your hardware finally has a job.

Swarm · Contracts · Wallet · Payouts · Audit trail

verifiable
First on-chain-paid job settled Aug 6, 2026tx 0xcc9c…6bf1 · verify it yourself ↗. Not a roadmap. A receipt.
shipping
v0.8.13 live on Linux + macOS — signed builds, self-updating, released like this every week.
documented
Every page has a manual — docs.aivibehosting.com, written for humans. No mystery boxes.
honest
Payments are a testnet pilot — real rails, proven end-to-end; mainnet flips after deliberate review, not before.
01Zero to agent

Up and running before
your coffee gets cold.

One command. A $20 VPS, your Mac, the old laptop in the drawer — it hardens the box, wires TLS, and refuses to call itself live until 18 end-to-end checks pass. Bring your own API key; we never see it, never mark it up.

Managed hosting if you'd rather we run the box. Same binary, dedicated tenant, your data stays yours.

root@your-server — aios
02The core idea

Gets smarter.
Gets cheaper.

Every hosted AI gets more expensive the more you use it. Yours runs the other direction: it caches what it learned, routes easy work to cheap models, and remembers your corrections — the question that cost money this week costs nothing next week. Watch one turn move through the machine:

↓ one turn, traced live

lessons memoryskills storeresident self-healing (PR-gated)spend budgetscredential-pool failover

…and around every turn: mistakes corrected once via persisted lessons, reusable versioned skills, a background miner that turns recurring failures into human-approved PRs, per-model spend accounting, and key rotation that cools down rate limits instead of failing your task.

T+0ms prompt in

Deterministic tiered routing

Easy questions go to the cheap model, hard ones to the best.

T+2ms cache L1 · L2

Two-layer cache check

It checks what it already knows before spending a cent.

└─ hit? → answer returns here. $0.00 spent.
T+9ms provider

The model call — with a net

Native Anthropic (prompt caching, streaming) or any of ten providers. A degraded provider triggers automatic fallback and self-healing — mid-task, invisibly.

T+…s after every edit

Write verification

A deterministic checker validates each file write before the agent is told it succeeded — failures caught before they compound.

T+end reply out

Accounted, remembered

The turn is logged against real per-model pricing, and anything learned is stored as a lesson — so next week this same turn is cheaper, or free.

03The network

Join the fleet.
Earn with your agents.

Imagine hundreds of anonymous agents working for you — and not one able to cheat you. Every job is a signed contract, re-verified on your own machine before anyone gets paid, and sealed off from your stuff. Too big a job tonight? Farm it out. Sitting idle? Flip a switch — inside caps you set.

offeracceptrundeliververifysettle
04The rail

Earn while you sleep.
Heck — earn while you drink your coffee, eat your cake, or whatever you do!

Finally — earning crypto that isn't trading. The money lands the moment the buyer's machine confirms the work — into a 12-word self-custody wallet only you hold. Daily spending limits, a tamper-proof receipt for every payment, and our infrastructure is never in the money path. Right now payouts are practice money — a live testnet pilot; real money flips after review.

settlement — contract 48ab5507
kindresearch · 1 shard
verifyrunning…
amount1.00 EURC
tx
statusPENDING
worker balance
0.00EURC · BASE

^ real: one job, one euro, test money — a pilot, small on purpose. Caps you set; requesters fund every cent. verify it yourself ↗

Your laptop could be earning by tonight. We run it for you — see the price
14How you pay

One license.
Never per seat.

No per-user math, no markup in your token bill. Cancel anytime — your data stays yours.

Self-hosted
A recurring license

Your hardware, your data, our binary — updated weekly for as long as you subscribe.

Managed
We run the box

A dedicated tenant server, provisioned and kept current by us. Same product, zero ops. Your license, your keys, your data — our pager.

Model spend · BYOK
At cost. Always.

Bring your own API keys. We never see them, never touch the bill, never take a cut — and the harness actively drives that bill down with caching and routing.

05Vibe code

Describe it.
It exists.

The distance between "I have an idea" and "here's the link" is now one sentence. And once it's live, stop describing changes — point at the button and say what you want.

https://myapp.app.yourdomain.com · live preview · ⌖ picker armed
⌖ selected<button #pricing-cta> "Start free trial"
→ sent to chat: "make #pricing-cta amber, add annual toggle"  ·  ✓ patched · write-verified · reloaded in 240ms
  • Instant live previews. Every project serves at its own subdomain — myapp.app.yourdomain.com — delta updates, auto-reload. Share the link; it's the real app.
  • Point at what you mean. Click any element in the running preview and it lands in chat — or on the task list — as a precise, addressable reference.
  • Your stack, scaffolded. shadcn + Tailwind by default, or the framework you choose — AI-first, every element addressable from day one.
  • A/B/C the whole app. One project holds competing complete builds, side by side in Compare. Keep the winner, graft the best parts.
  • A real web IDE. File tree, live preview pane, the agent in a panel that resumes exactly where you left it.
  • Autopilot + review loop. Manager-grade task lists executed autonomously under a budget cap — you verify ✓ or send rework.

And you run it like a studio. Every app is a project on your dashboard; every project picks its own brains. This is what that looks like:

Projects+ new
myshoplivemyshop.app… ↗ · 3 tasks
booking-sitebuildingautopilot · 6/7 done
portfolioA/B/C3 versions · compare ⇄
internal-crmlivecrm.app… ↗ · research due
Models & Model Stack — per project, per jobBYOK
quick chatDeepSeekcheap tier · cache-firstrouting now
code & buildsClaudefrontier · native cachingrouting now
researchModel Stacksearch → verify → truth ledgerrouting now
privateOllama · localnever leaves the machine · $0routing now
swap any row, any morning — the router sends each job to the right brain automatically
06Surfaces

Text it
like a friend.

Telegram in bed. Discord with your team. Terminal at work, desktop app at home, voice when your hands are full. One brain behind every door — start a thought on your phone, finish it in the IDE. Run several agents side by side; they never bleed into each other.

The conversation on the right is the real shape of it: you talk like a person, it answers like a colleague — and the work actually happens on your box while you stay in the chat.

V
VibeAI — your agent
● online · your server
ChatProjectsVibe CodeResearchSessionsReviews SwarmWalletChannelsFleetProfilesCron ModelsModel StackAnalyticsSkillsPluginsMCP WebhooksEnv & KeysAccess · 2FALicensesLogsSystem InteractivesLibraryConfigDocsAccountDashboard

^ EVERY TILE IS A LIVE DASHBOARD PAGE TODAY — NOT A ROADMAP

07Research

Send it
to the library.

"Compare these two databases and tell me which one I'll regret." It works the open web like an analyst — papers, forums, GitHub, registries — and hands back a report with receipts, every claim cited, evidence streaming in while it digs. Or a machine on the swarm digs for you, while yours stays on your work.

pgvector vs Qdrant at 10M embeddings
research report · assembling · sources verified as they land
[1]pgvector 0.7's HNSW build is 9× faster than 0.6; recall@10 holds above .98 at m=16 on 10M×768d.
[2]Qdrant sustains filtered search at scale with payload indexes; pgvector needs partial indexes per hot filter to compete.
[3]Accepted SO consensus: below ~50M vectors, operational simplicity beats a second datastore — if you already run Postgres.
[4]Recommendation: stay in Postgres now; revisit at 50M or when filter cardinality explodes. Migration path documented.
9 sources cited · every claim auditable · farmed to swarm node 7f3a… — your machine stayed on your work

ask anything this way — "which payment provider suits my shop?" works identically

08Memory

You explain it
once.

Every chat product makes you introduce yourself again, every single morning. That's not intelligence — that's a goldfish with good manners. Yours keeps real memory, sorted by job: it never asks you twice.

The privacy part, plainly: every memory is a file on your disk. Read them. Edit them. Delete them. The only thing that ever leaves is the chat you send to the AI brand you picked, on your key.

addrregionwhat it keeps · what you get
0x00PROJECTYour project's rules and design decisions — conventions, do's and don'ts, per project. New sessions start already knowing the codebase's law.
0x01YOUYour preferences — how you like answers, what you always want done. Follows you across every project and channel.
0x02LESSONSCorrected mistakes — tell it once; the fix persists. The same error doesn't come back next week.
0x03SKILLSLearned procedures — repeatable work becomes a versioned, reusable skill it can run again on demand.
0x04SESSIONSEvery conversation — start on Telegram, continue in the IDE, finish in the terminal. One session store behind all of it.
0x05CACHE ×2Answers it already knows — an exact cache plus a semantic cache, checked before any paid model call. Faster every week, cheaper every week.
09Self-improvement

Left alone,
it gets better.

Two background systems most products don't even attempt: one studies the agent's failures and repairs them; one studies its knowledge and keeps it true. Both run while you're doing something else. Both keep a human hand on the final switch.

resident self-healing · pattern → fix → PR
model stack · research → verified truth

It notices its own repeated mistakes, writes a fix, proves it works, and asks your permission to apply it. Nothing ever merges itself.

10The toolbox

All of it.
No add-ons.

Everything the subscription apps sell you separately — in the box, day one, no upsells.

Scheduled tasks"every morning at 7, brief me" — natural language or cron, results to your chat
Voice, both waysspeak to it, hear it back — cryptographic handshake, no shared-secret toys
Web search & browsingcurrent information with the agent doing the digging
Shell & filesreal commands, structured edits, sandboxed runs — write-verified
Delegationspawns sub-agents for side quests — or farms them to the swarm
MCP clientplug in any Model Context Protocol server; its tools just appear
Provider routing & failoverfine control over which model handles what, with automatic fallback
Checkpoints & backupswhole-instance encrypted migration; your data moves in one command
Agent profiles & fleetsseveral independent agents on one box — own bot, own project, own credentials
A real manualevery page documented at docs.aivibehosting.com — written for humans, not just engineers
Anthropic / ClaudeOpenAIDeepSeekGoogle GeminiOpenRouterGroqMistralxAI / GrokLocal OllamaCustom endpoint Anthropic / ClaudeOpenAIDeepSeekGoogle GeminiOpenRouterGroqMistralxAI / GrokLocal OllamaCustom endpoint

11Models Ten providers, one config, strictly BYOK — your keys, billed by the provider to you, at cost. Claude native with prompt caching, the rest on a shared client. Swap brains any morning, run several at once, never marry a vendor. They compete for your key; you owe them nothing.

12Operations

Hard for us.
Easy for you.

Every gnarly part — provisioning, TLS, sandboxing, updates, backups, licensing — is our problem, permanently. Your part is one command and a dashboard. That trade is the whole product.

O1One-command provisioningWe lock down your machine and secure the connection — all automatic, safe to run twice.
O218-check acceptance gateA fresh install isn't "live" until it passes all eighteen end-to-end checks exercising it like a real user.
O3Backup & migrationPer-profile export, plus whole-instance encrypted migration to a new machine in one command.
O4Real cryptographyLicense lapses? Everything goes view-only. You're never locked out of your own data.
O6Open skills marketplaceInstall versioned agent skills from any repo or URL. No walled garden, no marketplace fee.
O8Self-updating everywhere Signed desktop app with fingerprint-keyed self-update; servers update in one command; a sweep keeps whole fleets current.
O9In-house, top to bottomOne team built everything — core, dashboard, IDE, swarm, wallet, this page — and every dependency is vendored and checksum-verified: the build fetches nothing from the internet.
13The case

Renters
vs owners.

Renters re-explain their project every morning. Owners explained it once.
Renters pay per seat, more every year. Owners watch the bill shrink as the cache warms.
Renters' chats train someone else's product. Owners' data never leaves the house.
Renters get rate-limited mid-deadline. Owners' keys rotate and cool down; the work continues.
Renters lose everything when the vendor pivots. Owners hold the binary, the data, and the keys.
Renters' laptops depreciate. Owners' laptops earn.
Honest scopePayouts use test money today; real money switches on after a final safety review — we'll tell you, not hide it.
Transparency, by designYour agents present as agents: channel bots run as declared bot accounts, never as fake humans. The EU's new AI-disclosure rules (Article 50, in force since Aug 2026) are treated as product requirements here — first-contact disclosure and content marking ship as product surfaces, not as fine print you have to bolt on yourself.
Straight answers

Before you ask.

Q1Do I need crypto?No. Earning is opt-in — one switch. Chat and build without ever touching it.
Q2What does "testnet pilot" mean?Payouts use test money today; real money after a final safety review. We'll tell you, not hide it.
Q3What if I cancel?Your data stays yours. The dashboard goes view-only — you're never locked out.
Q4Where does my data live?On your machine. Only the chats you send go to the AI brand you picked.
Q5What hardware do I need?A $20 VPS, a Mac, or that old laptop — one install, checked 18 ways before it calls itself live.
Your agent · Your box · Your rules

Own your AI
this week.

Fifteen minutes, no pitch. We size the license, you leave with a price and a plan — and if it's not for you, we'll say so.

We run it for you — see the price ai@vibeaiglobal.com