exe.dev
Computers for developers and agents — spin up real Linux VMs so agents run in their own instances, not on your laptop’s last free GB of RAM
← AI Agents · Tech · Eve / Ian Borders · OpenClaw · Codex engineering · Multi-agent harness · Gym hack
Why this page exists
Friend and builder Ian Borders (Eve Studio / Mac control surface for agents — see /eve) flagged exe.dev as something to get excited about: spin up virtual machines so agents can run in different instances — isolated, durable, not fighting your local machine.
That matches a pattern we keep hitting: local agents eat RAM, mix credentials, and make every experiment feel risky. exe.dev is the “give the agent a real computer on the internet” layer. This page is field notes to learn and use it more.
One-sentence TL;DR
exe.dev is a subscription for fast, durable Linux VMs (KVM isolation, root, SSH, persistent disk, public HTTPS hostname) aimed at developers and AI agents — buy a pool of CPU/RAM/disk, spin up many VMs, pre-install coding agents (Claude, Codex, pi, Shelley), and treat each box as a sandbox or long-lived workspace.
What exe.dev is
Tagline from the product: Computers for developers and agents. Durable sandboxes that are fast, secure, and sharable.
Just a computer
Real Linux VMs — not a proprietary container shape. Root, apt, systemd,
SSH. Disks persist; stopped VMs cost disk-only in the model they describe.
Pool, not per-VM tax
You buy a pool of CPU, RAM, and disk. Spin up as many VMs as your plan allows (e.g. Personal: many VMs sharing the pool). Make a bunch.
Agent-native
New VMs can ship with coding agents pre-installed (claude, codex,
pi) plus web agent Shelley. Drive boxes over SSH or HTTPS from code.
Networked & shareable
Public hostname / HTTPS proxy with sensible auth defaults — share a web server like a Doc, keep private by default, IAM for sharing.
Core entry: ssh exe.dev (see
exe.dev).
Docs stress real isolation (KVM VMs, not shared kernels) and
minimal blast radius: the VM only has what you put on it, so a runaway agent
has less to exfiltrate than your full laptop.
Why Ian’s “spin up VMs for agents” pitch lands
| Problem on one Mac | exe.dev-shaped fix |
|---|---|
| One agent fills RAM; meetings + agents fight for CPU | Agent lives on its own VM; laptop stays for humans |
| Agent needs root / installs / long builds | Full Linux with root and persistent disk |
| Parallel experiments step on each other | Many VMs — one per task, brand, or agent role |
| Security: tools + secrets on the same machine as personal data | Put only task data in the VM; pragmatic sandbox while “agent security” is unsolved |
| Show a prototype without deploy theater | HTTPS hostname on the VM; share the running app |
Community language matches Ian’s excitement: people spin disposable VMs multiple times a day, hand agents micro-VMs for tests, or run multi-hour dedicated sessions — faster loop than waiting on full CI or thrashing a shared laptop.
Running agents on exe.dev
From their agent use-case docs: create a VM (e.g. ssh exe.dev new), then use preinstalled
agents for research, prototypes, installs, and long-lived work. Long-lived boxes need explicit
agent upgrades (image ships versions at build time; apt upgrade won’t move Claude/Codex).
- Sandbox posture: “Put your agent in a VM and let it be” — install the agent, give it a job, supervise lightly; isolation is the pragmatic control plane.
- Shelley: built-in web coding agent on the platform (they dogfood building exe on exe VMs).
- OAuth for agents: product direction is “agents should access what they need” without parking permanent leakable credentials on the box (blog).
- Agent skill: teach other coding agents about exe.dev via their skill docs / SKILL.md pattern.
Pricing snapshot (check live — moves)
Source: exe.dev/pricing (captured Aug 2026; confirm before budgeting).
| Plan | Rough shape |
|---|---|
| Personal | ~$20/month — e.g. 50 VMs, pooled disk (~100 GB), data transfer cap; VMs share pool resources |
| Team | ~$25/user/month — similar VM counts per user, admin/SSO, burst across teammates |
| Reserved Cloud Pool | Hourly large pool (thousands of VMs class) — serious agent/factory scale |
Default disk sizes and overage rates are documented on the pricing page. Shelley tokens often track underlying model provider rates (no markup claim — verify in billing).
How it maps to our stack
| Layer | Question | exe.dev role |
|---|---|---|
| Local / Mac agent | Always-on laptop agents (OpenClaw, Codex CLI) | Offload heavy or risky jobs to a VM; keep local for thin control |
| OpenClaw / Mac Mini (/openclaw) | Dedicated home server for one stack | Cloud complement: more instances without buying more hardware |
| Eve / Ian Borders (/eve) | Mac UI to run/chat/deploy agents | exe.dev = where agent compute lives; Eve-class tools = human front door |
| PowerLobster / harness | Tasks, squads, waves | Each task or agent identity can target a VM (or pool) as execution environment |
| Kitesurf (/kitesurf) | Agent browser on Cloudflare | Browse layer ≠ full OS; exe.dev is for install/build/long process |
| Safety (/gymhack) | Goal-seeking agents break real systems | Contain tools and credentials per VM; least privilege by machine, not hope |
Design rule
One sensitive job family → one VM (or class of VMs). Don’t give the “book a gym class” agent the same box as bank cookies and production admin. Isolation is product architecture, not just infrastructure fashion.
Practical pilot (learn by doing)
- Create an account on
exe.dev
and open the CLI path (
ssh exe.dev/new). - Spin one disposable VM for a single coding agent task (clone a repo, run tests, tear down).
- Spin a second long-lived VM for a recurring agent (heartbeat, weekly research) — upgrade agents explicitly.
- Map to a real product job — e.g. a product brain research agent for Para Living that only has market data tools, not full theflysales.com admin.
- Wire identity lightly — prefer short-lived OAuth / integrations over permanent secrets on disk (see their OAuth-for-agents post).
- Document cost — pool RAM/CPU vs number of always-on agents; stop idle VMs when disk-only is enough.
Official links
Related on this site
- Eve — Ian Borders / Mac agent control surface (why this tip landed)
- OpenClaw · architecture — local always-on agents
- Codex engineering · Codex Fast vs Medium
- Multi-agent harness map
- Kitesurf — agent browser (complement, not replacement)
- Gym hack · HF × OpenAI — why isolation matters
- God agent vs team
Field notes · August 2026 · Credit: Ian Borders for the nudge · exe.dev
Comments
Approved comments appear below. Log in once with GFAVIP — it applies across the whole site. GFAVIP login
View comments archive