Measured on the live system.

The health check runs on the same server as the API, so it shows that the service answered, not that every network path to it was open. Most of the traffic in this period was our own testing and the public demo. Availability targets and the SLA roadmap are on the Security page.

The home screen was showing during the read calls. A busier screen has a larger tree and takes longer to read. For an agent loop, ask for JPEG or WebP at reduced size: the full-size PNG is there for pixel-exact checks.

From the production log of the last 24 hours before the measurement, timed on the server: 1,424 taps with a median of 19 ms, 576 swipes with a median of 340 ms (the gesture itself lasts 300 ms), and 524 app launches with a median of 346 ms.

• Each phone ran 40 rounds of screenshot, foreground read and key press. In total 520 calls ran in parallel and all 520 returned 200.
• The slowest single call in the parallel run took 429 ms. The slowest in the one-at-a-time run took 337 ms.
• While the five ran, we opened Settings on one phone only and read the foreground app of all five at the same moment. One phone reported Settings. The other four still reported their home screen.
• Every screenshot came back at its own phone's size: 720×1560 from the four phones and 924×1480 from the tablet.
• The server has 8 CPU cores and 32 GB of memory and runs 8 devices. Each device has its own limit of 2 CPU cores and 3 GB of memory.

We drove five of the eight devices. The other three are kept for calls and for the public demo, so they were left out. The run above was paced. Since 10 October the API limits requests per API key and per phone (60 a second for each phone, bursts to 120), so phones driven from one address no longer share one budget: the same test run flat out, with no pause, returned 200 on all 390 parallel calls.

• On 10 October 2026, between 12:55 and 13:27 UTC, the container of every one of the 8 devices was replaced with a new one. Each new container was attached to the phone's existing storage volume.
• The storage volumes were created between 2 and 9 October. None was recreated.
• Apps installed before the replacement were still installed after it. One phone still had apps first installed on 2 and 3 October. The tablet still had apps first installed on 5 and 6 October.
• A phone with Google Play still listed its signed-in accounts after the replacement. We checked that the account entries were present. We did not open each app to confirm its session.
• There is no session to open or close in the API. A key can stop calling for a day and find the phone as it left it.

How it was taken: container and volume creation times from the container runtime, and each app's first-install time read from the phone, on 10 October 2026 at 18:21 UTC. There is no backup job for phone storage; see what we do not do yet.

Each account has its own private network, with connections between phones on the same network switched off. The phones are containers on one server and share its kernel; that limit is stated on the Security page.

Apps that require hardware-backed attestation or a certified device may refuse to run. That is the same on every plan; the pricing FAQ says more.

• pip install erius-phone-mcp into a new virtual environment finished in 12.5 seconds without errors.
• The MCP server started over stdio and listed 21 tools. Reading the device list, the screen tree, the foreground app, the installed apps and a screenshot all succeeded against a production phone.
• A Python script that installs an APK, launches it, reads the screen and saves a screenshot ran end to end. The install took 569 ms for a 5 MB APK.
• A Node script ran the same four-step flow on two devices at once in 2.4 seconds, and on a third account's phone in 2.6 seconds.

Both scripts are in the integration examples. The MCP server is on PyPI and on GitHub.

• Every number on this page was taken from the production system. None comes from a lab copy.
• The test machine sits about 2 ms from the API. Your own network round trip comes on top of each time.
• The phones tested were our own and those of our internal test accounts. No customer phone was used.
• ERIUS PHONE is in early access and runs on one server. These are measurements from one period on that setup.
• We will re-run the tests and change the date at the top of the page when we do.