Your laptop
Your files, your accounts, your home network. Microsoft says OpenClaw shouldn't run on a normal workstation, only in a dedicated VM or on a separate device.
Pre-seed · October 2026
Your agent's own computer, in the cloud.
Cubicles gives personal AI agents an always-on cloud desktop. It's separate from your laptop, you can watch it from anywhere, and it costs one flat price a month.
- Raising
- $750k
- Post-money cap
- $7.5M
- To build
- Our own servers
inbox-agent · Working
4 vCPU · 8 GB · Frankfurt
Chrome · Gmail
- opened Gmail in Chrome
- archived 38 newsletters
- drafted a reply to the supplier
- saved 12 suppliers to Notion
- ran the test suite · 42 passed
- messaged you: 2 emails need you
The problem
Personal agents need a computer. Every place they run today is the wrong one.
A Mac mini
$599 upfront. The base model sold out in the US in 2026, with scalpers near double retail. It still sits on your Wi‑Fi.
A headless VPS
No longer cheap: 8 vCPU / 16 GB on Hetzner now costs about $83 a month. And there's no desktop, so the agent can't use real apps and there's nothing for you to watch.
A dev sandbox
Expires in hours and forgets every login. Left on all month at 4 vCPU / 8 GB, it costs about $240 or more.
341 of 2,857
skills that Koi Security reviewed on OpenClaw's ClawHub marketplace were malicious. Owners want their agent in a separate room they can see into.
Sources: TechRadar (Microsoft guidance), Decrypt, Gizmodo, Pinggy, Hetzner pricing, Morph / Koi Security.
The solution
A full computer in the cloud that belongs to your agent.
- A real desktop. Browser, terminal, VS Code and any app you install.
- Always on, always signed in. It keeps its sessions for weeks, not hours.
- Walled off. It isn't your laptop and it isn't on your network.
- Watch it work. A live view of the desktop in any browser, including your phone.
- One flat bill. From $29 a month. No per-second meter.
The agents come installed: OpenClaw, Hermes, Claude Code, Codex and the ChatGPT desktop app.
research-agentWorking
Reading 12 supplier sites
Standard · 4 vCPU · 8 GB · 100 GB · up 23 days
code-agentWorking
42 tests passed
Pro · 8 vCPU · 16 GB · 200 GB · up 9 days
client-acmeAsleep
Files and logins kept
Standard · 4 vCPU · 8 GB · 100 GB
Why now
Models can use computers now. The bottleneck has moved to where they run.
81.8%
Claude Opus 5.5 on OSWorld 2.0, Anthropic's partial-credit launch figure. Claude was under 15% on OSWorld when computer use launched in October 2024.
323k
GitHub stars on OpenClaw, the personal always-on agent, now backed by OpenAI. Its owners are shopping for somewhere to run it.
Sold out
The $599 Mac mini in the US and China in 2026. People are buying hardware just to give an agent its own machine.
Long-running agents fail on state, not intelligence. They need a machine that stays on, keeps its logins and can be watched.
Sources: Anthropic system card, Wikipedia, Decrypt, SCMP.
Where we are
The product is built and live.
Capacity is the limit.
Live today
- Sign up with email, Google or Apple, then pick a plan
- Create an agent and pick a desktop plan; a fresh desktop is provisioned for it, agents pre-installed
- Live stream to your screen at up to 60 fps
- Share the cubicle to watch, or to use it together
- Dashboard, checkout and self-serve billing
- Automatic CPU caps on suspected mining
Plans
Most popular
Standard
$29/mo
4 vCPU · 8 GB RAM · 100 GB
Agents juggling browser, docs and sheets.
Pro
$55/mo
8 vCPU · 16 GB RAM · 200 GB
Coding agents, builds and data tools.
Studio
$79/mo
12 vCPU · 24 GB RAM · 300 GB
Several agents, heavy builds and media work.
Every plan: up to 1 Gbps, unmetered bandwidth.
Against the owner's real options
Nothing to buy, nothing on your network, and you can see it.
| Cubicles | Mac mini | Headless VPS | Dev sandbox | |
|---|---|---|---|---|
| A real desktop your agent can see | Yes | Yes | NoCan't use apps | Yes |
| Stays on and signed in | Yes | Yes | Yes | NoExpires in hours |
| Off your home network | Yes | NoSits on your Wi‑Fi | Yes | Yes |
| Watch it from your phone | Yes | PartlyNeeds remote setup | NoNothing to watch | Yes |
| Ready in a minute, nothing to buy | Yes | No$599 and weeks to ship | Yes | PartlySet up in code |
| Price | $29a month, flat | $599upfront, ≈ $25–29/mo over 3 years | $42+a month for 8 GB RAM; $83 for 16 GB | ≈ $240+a month if left on |
Estimates as of Oct 2026 from vendor pricing (Hetzner, DigitalOcean) and Gizmodo's Mac mini vs VPS cost analysis.
Market
A fast-growing category, and a beachhead nobody is serving.
AI agents market
$7.8B → $52.6B
2025 to 2030. A ceiling for agent software in general, not our TAM.
Budget agents can take over
$1.8B
UiPath's ARR alone. RPA customers pay $215–$1,380 per bot per month.
Our beachhead
Agent owners
Indie developers, founders and power users running OpenClaw, Hermes or Claude computer use. They already pay $5–$30 a month.
Bottom-up, illustrative
1% of OpenClaw's 323k stargazers at $29/mo≈ $1.1M ARR
5%≈ $5.6M ARR
10%≈ $11.2M ARR
Sources: MarketsandMarkets, The Investing Show, BrowserCat, Decrypt. Bottom-up rows are arithmetic, not a forecast.
Competition
Everyone else sells to the agent's builder. We sell to its owner.
E2B, Daytona and Cua bill per second for short-lived sandboxes. Windows 365 and AWS WorkSpaces sell to enterprises.
E2B
~$392
Windows 365
~$297
Cua
~$261
Daytona
~$242
DigitalOceanheadless
$48
Cubicles
$29
730 hours at each vendor's published rates. E2B includes its $150 Pro fee.
Business model
A flat price only keeps its margin on hardware we own.
$29 for 4 vCPU / 8 GB is about 40% less than DigitalOcean's $48. Hetzner raised prices 2.1–2.75× in June 2026.
One server, our model
| Hardware | 2 × EPYC 7H12 · 1 TB DDR4 · 15 TB |
|---|---|
| Cost | ≈ $18,000 |
| Desktops | ~180 × 8 GB (1.5× RAM, 2.8× CPU) |
| Hardware over 36 months | $500 / month |
| Colocation, power, bandwidth | $399 / month |
| Card fees (~4%) | $210 / month |
| Cost per desktop | ≈ $6 / month |
Per server, per month
$5,220Revenue with all 180 slots sold at $29 per 8 GB slot (the Standard price)
~79%Gross margin with every slot sold
~71%Gross margin with 70% of slots sold
4–6 moHardware payback per server, from fully sold to 70% sold
The ask
$750k post-money SAFE, $7.5M cap
Pre-seed to take Cubicles from one server to our own fleet, and expand our team and services.
- Servers
- 1 → 4
- Desktop capacity
- ~720
- Runway
- 24 months
- 4 servers$72k10%
- Network, storage, spares$18k2%
- Colocation and bandwidth, 24 months$33k5%
- Team: engineering and marketing. 24 months$400k53%
- Go-to-market: guides, launches, community$150k20%
- Legal, tools, buffer$77k10%
Four servers full: ~$20.9k MRR at ~79% gross margin.
What the round buys
Twenty-four months. Servers follow demand.
-
Months 0–6
One server, first customers
Rack our first server, make our first infrastructure hire, add Windows desktops, and put the effort into marketing and signing up paying customers. Study the market as we go: who needs Cubicles most, which use cases and plans actually sell, and where demand is strongest. Start partnerships with universities and with developer and agent communities.
-
Months 6–12
Proven demand, new regions
Put our spending into the customers and plans that sold in the first six months. Expand those partnerships: universities, so students and research labs run their agents on Cubicles, and agent communities like OpenClaw and Hermes, so we're where new owners look for a place to run. When a server nears full, we rack the next one, placed in the region where demand is coming from, working toward the EU, the US and East Asia.
-
Months 12–24
Grow with demand
Keep adding servers the same way: one at a time, only when the last is near full, so every server is profitable before we pay for the next. Rack our first GPU server to test demand for private GPU compute, the next step of the plan.
Where this goes
We're building an AI hardware company. Agent desktops and harnesses are step one.
-
Step 1 · Now, this round
Desktops and harnesses for AI agents
Always-on virtual desktops for agents to run on, plus the AI harness that runs the agent, set up and kept running for you. One flat price.
- Runs on
- CPU and RAM servers
- We sell
- Desktops and harnesses, by the month
-
Step 2 · Next
Private GPU compute
GPUs to run private open-source LLMs of any size, from one card to a full node. The model sits next to the agent, and nobody else sees the prompts.
Why next: the most asked question in OpenClaw and Hermes communities was whether we'd let them run their own open-source models.
- Runs on
- GPU servers
- We sell
- Dedicated GPU capacity
-
Step 3 · Then
Our own models, sold as tokens
We run open-source LLMs on our own fleet and sell tokens through an API. Spare GPU hours from step 2 become token revenue.
- Runs on
- Our GPU fleet
- We sell
- Tokens
Each step runs on what the last one built: our own racks, colocation and ops, and customers whose agents already live with us.
Every agent owner should have a computer they can see into.
Raising $750k on a $7.5M post-money SAFE to build the servers that get us there.
Sources
© 2026 Cubicles · cubicles.cloud · Confidential