A colleague does the process once, explaining as they go. Arpa compiles the recording into a reliable workflow and runs it on a fleet of computer-use agents, inside the apps you already have. No APIs, no integration project, people in the loop.
items passed across three legacy-app scenarios, checked against each app's own data
model cost per item (€0.004 to €0.017 measured), including the AI reviewer that checks the work
of steps replay deterministically, in milliseconds, without calling a model
passwords or one-time codes ever shown to a model. Secrets are typed by the worker and masked in every frame
Measured on our Legacy Lab: deliberately awkward web apps with pop-ups, expiring sessions, iframes, 2FA and SSO.
Built for the software nobody wants to integrate with
Traditional RPA needs developers to script every click. Pure AI agents improvise every click, slowly and expensively. Arpa learns the process once from the people who know it, then runs it the cheap way, with AI only where it adds something.
The desktop recorder captures the screen, every click and keystroke, the app's accessibility tree and the narration. Passwords and personal data are blurred on the device, before anything is uploaded.
macOS today · Windows in previewArpa infers the inputs, the decisions and the checks. Steps that never change become deterministic, judgement calls follow your policy documents, and sensitive steps wait for approval. No hand edits are needed.
recording → published process in minutesSandboxed workers run the queue in parallel. Each step tries the cheapest way first and escalates only when the screen surprises it, up to a person who can answer or take over the screen.
web today · Windows desktop in previewEach step carries a ladder of ways to locate its target: stable ids, labels, nearby text, a picture of the element, a description for the agent. After it acts, it checks the result: the right window, the right value, the right record.
Scripted bots break when a button moves. Agents that think about every click cost cents per click and still guess. Arpa replays what it knows in milliseconds and escalates one rung at a time. Every rescue becomes a proposed fix, so the workflow gets cheaper the more it runs.
The exact element that worked last time on this screen.
~ 0 · msStable ids, accessibility names, test ids, labels.
~ 0 · msRenamed ids and moved labels, scored for uniqueness.
~ 0 · msA small vision model points at the element on the screenshot.
~ €0.001A bounded computer-use agent recovers the step, within a budget.
centsAnswers a question or takes over the screen. The answer teaches the workflow.
your call€0.0106 per new hire, all model calls included, and the per-step cost keeps falling as successful recoveries are folded back into the workflow.
Start in practice mode, where workers do everything except the final save. Move to "ask before sensitive steps". Go fully autonomous when the numbers say so. Whatever you choose, an unsure worker asks instead of guessing.
Each run keeps a screenshot per step, how each element was found, what it cost, and an AI reviewer's findings, which compare the work item with what actually landed in the system. Postconditions catch mistakes as they happen.
Point Arpa at the AP mailbox and every e-mailed invoice becomes a work item, with its attachments stored as documents. Watch an SFTP folder, run on a schedule, take a signed webhook from your ERP, or upload a spreadsheet.
Every model call goes through one metered gateway. You see the cost per item and per process, which tier did the work, and the time handed back against a human baseline.
Every scenario runs on the platform as it ships: queue, fleet, approvals, review. The result is then reconciled against the target app's own database.
Expiring passwords, vendor lookups, GL coding from the AP policy, approvals above €10k, duplicates skipped.
Authenticator codes coordinated across parallel workers, a cross-site signing widget, invalid payments sent to a person.
Keycloak sign-in, a five-step wizard, a people picker, an ID document uploaded byte for byte, rejections routed to HR.
A narrated demo became a 15-step process with no hand edits. The fleet booked real invoices; an agent cleared a surprise dialog and proposed the fix.
Legacy Lab apps are built to fight automation: random password-expiry prompts, ASP.NET ids that change on every load, cross-origin iframes, idle timeouts, duplicate-detection dialogs. Every number here comes from a recorded run report.
GDPR and the EU AI Act shaped the architecture from the first line, not a compliance layer added later.
Credentials live in a vault and are sealed to one worker session. They are typed at keystroke level only into fields found by structure, and masked in every frame, log and recording.
The whole platform runs on Docker Compose on your own hardware: the same code as the hosted service, with your network, your vault and your model keys. Air-gapped mode is on the roadmap.
An append-only audit log, a screenshot per step, the model routing behind every decision, and human oversight on every sensitive step, ready for AI Act logging and oversight duties.
Recording is started by the employee, with a visible indicator. Recorders are pseudonymised, recordings never feed per-employee analytics, and redaction happens on the device.
Every tenant's rows are fenced off by Postgres row-level security, not by application code alone. Roles cover admins, process owners, supervisors, reviewers and auditors.
One model gateway meters every call and enforces budgets. Use the open models we provide, or bring your own keys for Anthropic, OpenAI or any compatible provider.
One make up and you have the console, the fleet, the vault, single sign-on and the model gateway. Workers connect outbound only, so no inbound firewall holes are needed.
The same platform, hosted and operated by us in European data centres, with sandboxed Windows and web workers on demand and EU-resident models.
Arpa is in private preview with a small number of European teams. Bring us the process everyone dreads, and we'll show you a worker doing it.