An open, headless simulator that coding agents can drive for Expo apps in CI

[Hansade2005]

Coding agents now build Expo apps, and they need somewhere to run and test them. Today that means a full Android emulator or iOS Simulator per agent, which is heavy when you want many agents, or many parallel CI jobs, on one machine.

I have been working on this problem in an open-source project, Shipvo (Apache-2.0), and I would like to share what we measured and ask what the Expo team and community would want from a standard here.

What we built: a headless simulator that runs the web build of an app in a phone-sized Chromium context, with an Android or iOS frame, and exposes it to agents over MCP (read the accessibility tree, tap, type, scroll, set device state such as permissions, connectivity and payment results).

What we measured, all published with scripts:

• 110 simulators open at once on a 2-CPU, 7.8 GB machine, each answering, at 57.1 MB each: https://github.com/shipvo/shipvo/blob/main/docs/product/sim-bench/scale-2026-11-07.md
• First tap about 10.7 s (Android) and 10.2 s (iOS) after the process starts, on a clean GitHub-hosted ubuntu-latest runner: https://github.com/shipvo/shipvo/actions/runs/42355471525
• Parity against the Android emulator and iOS Simulator on 13 screens of two purchase flows: https://shipvo.dev/docs/parity-results/

Questions for the community:

1. Would a documented, open interface for agent-driven app testing (accessibility tree plus actions over MCP) be useful in Expo's tooling?
2. Which device states matter most to test: permissions, connectivity, payments, notifications, deep links?
3. Where should simulator-versus-device parity be measured and published so teams can trust it?

Disclosure: I am the author of Shipvo. I am not asking you to adopt it. I would like the discussion to shape what a shared approach should look like, and I am glad to contribute the measurements and code.