The worktree shuffle
Get out of each other’s way.
Branches on one machine still collide on ports, node_modules and build caches. Each task gets its own files, processes, dependencies and ports.
One task · one microVM
Stop juggling worktrees and fighting over ports, caches and CPU on one machine. Each task gets an isolated Linux microVM that pauses after 15 idle minutes, even with your agent open, so the compute stops too.
✓5 free agent-hours✓paused time is free✓no card
Bring your own account. Codex connects once; Claude Code and Pi sign in once in each task and stay signed in across pause and resume. Model usage is separate.
Works with Codex, Claude Code, Pi and SSH.
Why MicroDev
Worktrees separate files. Your agents still share one machine: its ports, its caches, its CPU and its power switch. MicroDev gives each task an environment of its own.
The worktree shuffle
Branches on one machine still collide on ports, node_modules and build caches. Each task gets its own files, processes, dependencies and ports.
The shared-machine bottleneck
Agents on one laptop queue for the same cores. Give each agent its own Linux microVM and run builds and tests side by side.
The always-on compute bill
An agent waiting on you still holds a machine. A task pauses after 15 idle minutes, even with the agent open and your terminal or the Mac app attached: its compute is released and the workspace stays saved.
How it works
From a prompt to a pull request, then out of the way.
Start your first taskStart from any terminal (or, once it’s released, the Mac app). Pick the repository, the agent (Codex, Claude Code or Pi) and its prompt. Codex runs on the account you connect once; Claude Code and Pi sign in once in each task.
microdev start OWNER/REPO --agent codex --prompt "…" A fresh root filesystem boots, your repository is cloned and your agent starts on the prompt. Your Mac stays the control surface.
fresh root fs · repo cloned · agent started The task has its own guest kernel, files, processes and network boundary. Tasks from different users share hosts; each one’s VM, disks and network are kept apart from the others and from the host’s private services.
own kernel · own files · own network In the Mac app (coming soon): review the diff in place. Comment on a line and send it back to Codex. When it looks right, open a pull request with exactly the diff you reviewed.
review · comment · open pull request Pause saves the workspace and agent sign-in, sessions and settings, then compute stops. A task also pauses itself after 15 idle minutes, even with your agent open and your terminal or the Mac app attached; builds, tests, downloads, streaming model calls and typing count as busy.
microdev pause ENVIRONMENT_ID One agent, many microVMs
Give the coding agent on your laptop the whole job. It starts a MicroDev task for every piece, each in its own microVM with its own kernel and network. Once you’ve connected Git, the agent in each task pushes a branch with your credential and the agent on your laptop opens a pull request for each one.
The agent’s side
Any agent that can run a command can do this: the skill at https://microdev.dev/skill/SKILL.md teaches it to start, watch and pause tasks, each with its own agent and prompt: microdev start OWNER/REPO --agent codex --prompt "…" --json. Each task’s Codex works on the account you connected, with no sign-in. Claude Code and Pi tasks start the same way but need one sign-in per task (microdev attach ID).
Your side
Every task shows up in the CLI (and in the Mac app once it’s released) like any other. Review each pull request on GitHub. A task pauses itself after 15 idle minutes, or the agent pauses it.
Capacity
Concurrency follows your plan: 2 tasks at once on the trial, 2 per seat on Pro and 4 per seat on Team, so the agent starts the next task as one finishes.
Compute explorer
A machine left on keeps allocating compute between tasks. A MicroDev task pauses after 15 idle minutes. Try your own numbers.
An illustration, not a price quote. It compares allocated compute only and excludes subscriptions, storage and model costs. Running time includes waits for tools and models; paused time is excluded from usage.
See plans and pricing75%
less allocated compute in this example
Illustration At the Pro overage rate of $1.00 per agent-hour, that is $120 versus $480 a month. Your bill differs: it ignores included hours and the seat price.
The persistence contract
A completed pause saves your task. Resume builds a fresh environment around that progress. Save files before you pause.
The Mac app
Coming soon: start a task with Codex, Claude Code or Pi from the composer and watch it work in a native terminal. Comment on the diff line by line, send the comments back as one batch, verify the answer and publish the pull request. Nothing runs on your laptop.
An illustrated walkthrough with sample data, not a screen recording; a real task takes minutes. Batch feedback in the agent’s own conversation works with Codex. The Mac app in the docs
Three ways to start
Every path starts on the same free trial: 5 agent-hours over 14 days, up to 2 tasks at once, no card.
Terminal
Sign in with GitHub, then launch the sample repository in its own microVM.
curl -fsSL https://microdev.dev/install.sh -o install-microdev.sh
sh install-microdev.sh
export PATH="$HOME/.local/bin:$PATH"
microdev login
microdev demo
Your agent
Paste this line to your coding agent and let it take it from there.
curl -fsSL https://microdev.dev/skill/SKILL.md Browser
Sign in with GitHub, add an SSH key and launch the sample repository. Nothing to install first.
No. The trial is 5 agent-hours (300 minutes) of running time over 14 days, with up to 2 tasks at once, no limit per run and up to 10 new tasks an hour (resuming a paused task doesn't count against that). Paused time is free. The control plane enforces those limits, and nothing charges you automatically. Upgrading is your choice.
You bring your own agent account. The Codex, Claude Code and Pi CLIs are preinstalled. Codex runs on the account you connect once, with no further sign-in; Claude Code and Pi sign in once in each task and stay signed in across pause and resume. Model usage is billed by your model provider, separately from MicroDev compute.
Your whole workspace, byte for byte: every branch, stash and commit, and staged, uncommitted, untracked and ignored files. Codex, Claude Code and Pi sign-in, session history and settings, so a resumed agent starts signed in and past onboarding. Installed packages, other home-directory files, shell history, agent caches and logs, environment variables and running processes are not kept: resume boots a fresh copy of the pinned system image and runs your dotfiles again. See the exact contract.
On subscription plans the contract is backed by the pause guarantee: if a completed pause loses committed, uncommitted or untracked work in your workspace, that billing period’s subscription fee is credited against your next renewal (a month’s fee on monthly plans, the year’s on annual plans). How it works.
No. A build or test run spends CPU, a download or streaming model call moves data, and typing in the terminal is activity; any of them keeps the task running. Auto-stop needs 15 continuous minutes in which every minute stayed under 2% of one vCPU and moved 32 bytes a second of data or more for less than half of it (packet headers and your SSH sessions excluded), with no typing and no transcript change. An attached terminal or Mac app nobody types into doesn't count. A tool that only sleeps or polls now and then, with no CPU, is paused like a finished agent, with its workspace kept. On the trial, tasks also pause when the trial’s running time is used up or its 14-day window ends.
Each task runs in its own Firecracker microVM under the jailer: its own jail, cgroup, root and workspace disks, PID and mount namespaces, and a host UID and network namespace that no other running task shares. From inside the guest, even as root, other tasks’ VMs, the host’s own services, cloud instance metadata and private network ranges are unreachable; only the Internet is. A task placed in a slot another task left starts with none of its connections. Physical hosts are shared: there is no dedicated-host guarantee.
A worktree is another checkout, but its processes, ports and installed dependencies still live on the same machine. MicroDev gives each task a separate Linux microVM, its own workspace and its own compute lifecycle. You can still use worktrees inside a task.
Yes. Create an organization, invite your team and buy seats. Seats pool across the organization: each one adds tasks at once and included agent-hours. Plans are Pro $29, Team $59, Business $119 per seat per month; Team adds usage by member and starts at 2 seats. See plans and pricing.
A little space goes a long way
Try a real isolated environment. Your first 5 agent-hours are on us.
5 agent-hours · 14 days · no card · no automatic charge