warden leases iOS simulators, Android emulators and real phones to your coding agents, drives the app on them and keeps what they checked as tests. Devices, ports and builds, leased across every worktree on your Mac. One CLI, free and open source.
$ npm i -g @delacour/warden && warden installInstalls the CLI and the skill for every agent it finds.
warden batch salient-e2e37 Salient e2e specs · 5 iPhone 17 simulators · 3m25s (13m55s on one) · 37/37 passed · shown at 5× speed
Claude Code+hook
Codex+hook
Cursor
GitHub Copilot
Gemini CLI
OpenCode
Windsurf
ZedEvery agent the skills CLI supports gets the warden skill. Claude Code and Codex also get a SessionEnd hook that releases what the session leased.
Leasing, device control and e2e tests in one tool, on one driver, so what an agent checked by hand passes the same way as a spec.
One global CLI, then warden install adds the skill to every agent it finds. There is no daemon to start.
$ npm i -g @delacour/wardenAsk for a platform. You get a booted simulator no other agent, worktree or shell on the Mac will be handed.
$ warden claim ios --jsonRead the screen, tap and type with warden ui, then keep the same checks as a spec that warden e2e runs on many devices.
$ warden ui snapshot -iEach agent works in its own worktree. warden hands each one a simulator, a Metro port and the build for its native fingerprint, and never hands the same one to two owners.
Every agent, worktree, CI job and shell on the Mac claims from one sqlite registry. Each claim is a transaction, so two agents asking at the same instant are served in turn, never both.
A leased device belongs to one holder until it releases it or the TTL runs out. Wrapper commands heartbeat every 30 s and release when they exit.
Read more →Nothing free that fits? warden boots or creates one, up to what the Mac's RAM and CPU can hold, and shuts down idle ones first when it needs room.
Read more →Past capacity, claims wait their turn instead of failing, by --priority and age. --no-wait fails at once; warden queue shows where you are.
New iOS simulators clone from a settled golden image: 3.7 s clone plus 11.9 s boot, instead of 146–239 s for a fresh one.
Read more →A warden device nobody has leased for 20 min is shut down to free RAM. warden sims prune deletes long-unused ones to free disk.
A device that fails its health check at claim time is quarantined, and the claim tries another. Release one as bad yourself and it is set aside for 24 h.
Read more →Metro ports are leased like devices. App builds are cached by native fingerprint and shared by every worktree, so a native change builds once.
Read more →Claim a plugged-in Android phone or iPhone and drive it with warden ui or a spec. warden never boots or shuts one down, and only uninstalls what it installed.
Once it holds a simulator, your agent runs the app, drives it, reads what went wrong and keeps what it checked as a test. Every step stays on its own device.
Run it
# worktree 1: feature/checkout$ warden dev ios --json# worktree 2: fix/login-crash, at the same time$ warden dev ios --json# each gets its own simulator and Metro port,# and the cached build for its native fingerprintDrive it
# your agent, on its own simulator$ warden claim ios --json$ warden ui open com.example.app$ warden ui snapshot -i@e1 [textbox] "Add a todo" id=todo-input value="" enabled hittable@e2 [button] "Add" id=todo-add enabled hittable$ warden ui fill @e1 "Write e2e specs"$ warden ui press @e2 --settle$ warden ui wait text="Write e2e specs"Debug it
# what broke, from the running app$ warden dev logs --kind bundle_failed,error$ warden ui exceptions$ warden ui network --status 4xx# why it is slow$ warden perf record --all --duration 15s --udid $UDID# or all of it over MCP$ claude mcp add warden -- warden mcpKeep it
import { expect, test } from "@delacour/warden-e2e"; test("add a todo", async ({ screen }) => { await screen.getByTestId("todo-input").fill("Write e2e specs"); await screen.getByTestId("todo-add").tap(); await expect(screen.getByText("Write e2e specs")).toBeVisible(); await expect(screen.getByTestId("remaining")).toHaveText("2 left");});Run the suite
Take it offline, break one host, move the GPS, jump the clock or fail Face ID, from a spec or a terminal, on a leased simulator. The app runs unmodified, and warden undoes every condition when the lease ends.
warden ui biometrics fail and match answer the real system prompt. Face ID and Touch ID on simulators, fingerprint on emulators, enrolled with warden claim --biometrics.Read more →warden is advisory, not a sandbox. What it enforces is its own reach.
A device warden didn't create or lease is read-only: never handed to an agent, shut down, erased or installed on unless you pass --adopt.
warden ui refuses a device leased to someone else, and heartbeats yours on every call. warden check exits 2 when a device is someone else's.
No server to run and no open port: leases are rows in one sqlite file, and warden's helpers talk over unix sockets. The network proxy binds loopback, only while it shapes traffic.
Read more →Wrappers release when they exit. The Claude Code and Codex hook releases what a session leased and shuts down what it booted; device conditions are undone on release.
Read more →Paste this into Claude Code, Codex, Cursor or any agent in your app's directory. The warden skill tells it the rest.
Install warden (npm i -g @delacour/warden && warden install) and set this Expo app up with warden init. Run the app from this worktree with warden dev, then drive it with warden ui and check the home screen renders. Save what you checked as an e2e spec under e2e/, run it with warden e2e, and release the simulator when you're done.
The jobs agents get asked to do on a mobile app, and the warden command behind each.
Five agents, five worktrees, five simulators and five Metro ports, side by side on one Mac.
Read more →The agent changes a screen, opens it on its own simulator and checks it before it says done.
Read more →Walk the steps from the issue on a clean device and keep the repro as a spec that fails until the fix lands.
Read more →Offline, a failing host, an expired trial, a failed Face ID: the states nobody tests by hand, as specs.
Read more →Only the specs a diff reaches run, in CI or a pre-push hook, across leased simulators.
Read more →Record CPU, memory, frames and startup before and after a change, and diff the two.
Read more →Short answers. The docs have the long ones.
One open-source CLI that leases iOS simulators, Android emulators and phones to coding agents, drives the app on them, and runs e2e specs across many of them at once. Devices, Metro ports and app builds are leased per worktree, so agents in parallel never collide.
No. It is advisory: the skill tells agents to use only devices a lease hands them, and raw simctl or adb are not intercepted. What it enforces is its own reach: warden ui refuses a device leased to someone else, and warden only shuts down, erases or deletes devices it created.
Every lease has a TTL (30 min, 5 min for a bare agent warden claim) and lives only while it is renewed. warden run, dev, batch and e2e heartbeat every 30 s and release when they exit, so a killed wrapper frees its device within about 90 s; one that vanishes costs the device until the TTL runs out.
warden boots or creates another, up to a capacity it derives from the Mac's RAM and CPU. After that the claim waits in a queue ordered by priority and age, up to --wait (30 min by default); --no-wait fails at once.
Yes. Android emulators are leased like simulators, and warden ui and specs drive them. Plugged-in Android phones and iPhones are opt-in with --physical; device conditions and biometrics are simulator and emulator only. iOS needs macOS with Xcode.
Any agent the skills CLI installs into gets the warden skill: Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode, Windsurf, Zed and more. Claude Code and Codex also get the SessionEnd cleanup hook.
Yes, for debugging. warden mcp serves a warden dev session's Metro output, console, exceptions, network, eval, logpoints and profiles over stdio. Claiming devices and running specs stay on the CLI.
Builds, the dev client and Metro target Expo and React Native. warden ui and specs drive any installed app by bundle id, finding elements by testID, label, text or role.
Nothing. warden is MIT-licensed and runs on your own Mac or CI runner. No account, no cloud.
$ npm i -g @delacour/warden && warden install