AskUI Platform | One app: write, run, schedule, prove
One app. Write, run, schedule, prove.
Validation, documentation and operation are the same app. The CLI runs the same engine in CI.
A computer-use agent works your software the way a tester does.
No selector, no recording, no object map. For every step the agent repeats one cycle: which is why a suite survives the UI changes that break selector-based automation.
01 Read
Take a screenshot of the device and see what a tester would see.
02 Decide
Interpret the step against what is actually on screen right now.
03 Act
Click, type and scroll, using only the tools the project permits.
04 Verify
Compare the result against the step's expected result, and record it.
“The step's expected result is the contract; the path is only the execution log.”
If a person can reach it, the agent can drive it.
The agent works from the screen, so the target does not have to cooperate. It needs no API, no test hooks, and in the hardest case nothing installed on it at all.
Desktop
Windows, macOS, Linux and Windows Server, driven over gRPC.
Native apps, thick clients, anything with a window.
Nothing installed
HDMI capture plus USB or Bluetooth HID, so the target needs no agent at all.
HIL benches, head units, kiosks, sealed appliances.
Many machines
One agent driving several machines in step, in one flow.
A simulation environment and an operator HMI, together.
Machine HMIs
Line panels and industrial control screens, reached the way an operator reaches them.
Where no selector has ever existed.
Android
Real devices over adb, including device racks.
POS handhelds, scanners, phones.
iOS
Simulators over idb.
Companion apps beside the cockpit or the till.
Web
Any browser, through Playwright when structured signals are there.
Fast where the DOM helps, screen-driven where it does not.
Citrix and VDI
A remote session is just a screen, so the agent works inside it.
No instrumentation, nothing installed in the image.
Nobody logged in? Still runs.
AgentOS installs as a Windows SYSTEM service, so a machine with no session open is still a machine the agent can drive: it survives an RDP disconnect, reaches the logon screen, and works a background session overnight.
RDP disconnect recovery available
Login screen automation available
Ctrl+Alt+Del available
Background sessions available
Signed binaries · silent install available
Four components, one engine.
AskUI Desktop Author, run and review: projects, devices, reports.Docs AskUI CLI The same runs, headless, for CI and schedules.Docs AgentOS Gives the agent eyes and hands on a machine: screen, keyboard, mouse.Docs AskUI Hub Workspace, members, tokens and billing.Docs
The model behind the “decide” step runs through the AskUI hub by default, or through your own provider: Anthropic, any OpenAI-compatible endpoint, or self-hosted.
The whole job, in one window.
WriteRunProveConnect
The test is a Markdown file.
Steps in plain language, preconditions and postconditions. No selectors to maintain when the UI moves.
Your testware is a folder in your Git repo.
Every artifact is a plain file: cases, plans, keywords, the prompts that describe your system, the tooling the agent may use. Diff it, review it, branch it. Nothing important lives only in our database.
my-project/git
tests/
rules.md setup.md teardown.md
login_test.md
checkout/
pay_with_card.md
plans/ smoke.md
procedures/ login_to_ui.md
prompts/ report_format.md
utils/ tools.json mcp.json
devices.json schedules.json
tests/ Suites and cases. A file is a case; a folder is a suite with its own setup and teardown.
plans/ Test selection: what makes up the smoke set, the regression pack.
procedures/ Keywords. Write a sequence once, call it from any case with parameters.
prompts/ What you would tell a new tester on day one about the system.
utils/ The permitted tooling: built-in tools, MCP servers, custom C# tools.
devices.json The test environments.
schedules.json The execution schedule for recurring runs.
agent_workspace/ Evidence per run: logs, reports, screenshots. Not versioned.
Phases in, evidence out.
THE ORDER OF A RUN
- setup.md Entry criteria first. A failing setup marks the tests it guards as broken, so they never run against a bad state.
- the cases Each case is one autonomous agent execution, streaming into the live log.
- teardown.md Runs last, even after failures.
- rules.md Standing instructions that apply throughout, accumulating per level.
WHAT EVERY RUN LEAVES
A verdict PASSED · FAILED · WARN · SKIPPED · BROKEN
A screenshot per step, as the agent saw it
The transcript every tool call and decision
The report in agent_workspace, per run
Give the agent more hands.
SQL over ODBC, HTTP, shell and file I/O are built in. Add MCP servers or your own C# tools, and keep credentials out of the transcript.
One command in CI.
Gate a release, or schedule the nightly report.
ci · nightlytoken: release-train
$ askui run tests --plan nightly
214 tests · 4m 12s
exit 0 all executed tests passed
CI/CD integration Pipeline examples Scheduling
Four ways to run it. The same app in all four.
How much of the platform sits on your side is a deployment choice, not a different product. Nothing about authoring or running a task changes between them.
01 Hub
Sign in and go. Inference is routed and metered by us.
Hub + your model
Your Anthropic or OpenAI-compatible endpoint, our hub for workspace and billing.
03 Licence key
Validated offline. The app never contacts our hub at all.
04 On-premise
The hub inside your perimeter, with unmetered agents and hardware control.
Security, procurement and the paperwork
Next step
See it on your own systems.
We scope the trial with you and run it on your real environment. A person answers, within one business day.
Scoped with you first A short call about targets, scale and deployment.
Terms agreed before the start Scope and conversion terms in writing, up front.
Nothing leaves your perimeter On-premise and air-gapped deployment. ISO 27001, GDPR.