An Anthropic computer-use agent — Claude with screen and keyboard control — that drives your legacy GUI like a human. Logs in, navigates, enters data, saves, logs out, reports status. Runs on a dedicated VM, on a schedule or on triggers. The capability productized stacks structurally cannot deliver.
If your staff spends 10–20 hours a week clicking through ancient software — software with no API, a vendor that won't build one, and your only previous option being "hire more staff" — this is the build that breaks the constraint. It's the service that productized AI agencies structurally cannot deliver, and it's the highest-moat service in our catalog.
Practices on EHRs from the 2000s, behavioral health on systems that have been "moving to a new platform next year" for a decade. Hours of daily data entry that nobody can automate because the EHR has no integration story.
Old practice management software, vendor claims portals, state filing systems, county property records. Critical to the business, no API, vendor unmotivated to build one.
State licensing boards, county recording offices, federal compliance portals. Where the ergonomics are bad on purpose and your only choice has historically been "have a person do it."
Operations stuck on legacy software face an ugly tradeoff: tolerate the labor cost forever, attempt a multi-year platform migration, or live with the constraint and hire around it. Three reasons this build changes the math.
Adding a data-entry FTE costs $50–80k all-in. The role has high turnover. New hires take 2–3 months to become productive. The math gets worse, not better, every year.
Replacing the legacy system means data migration, retraining, vendor selection, regulatory re-approval (in healthcare). Most operations live with the legacy system for years past the point where it's clearly costing more than it's worth.
Repetitive, deterministic, precise — the work that traditionally went to RPA tools (UI Path, Automation Anywhere) but with brittleness that made those tools expensive to maintain. Computer-use agents, with vision-model resilience, change the economics.
Anthropic's computer-use API gives Claude screen-vision and mouse/keyboard control. Around that primitive, we build the production scaffolding that turns a research demo into a system you can rely on for daily operations.
"On this trigger, log into X, navigate to Y, fill in Z from this data, save, log out, report status." Each task defined in code with explicit success criteria. Triggers can be schedule-based, event-driven (a webhook, a row in a database), or manual.
The agent runs on a dedicated Linux VM with the legacy software installed. Headless when possible, GUI-rendered when required. Isolated, monitored, and recoverable.
Every step of every task captures a screenshot. Full visual audit trail. When something goes wrong (and at the start, things will), you can see exactly what the agent saw and decided. Critical for regulated workflows.
Modal dialogs, login timeouts, network blips, transient slowness — all the failure modes a human handles instinctively. The agent retries with backoff, escalates when stuck, never silently corrupts data.
When the agent encounters something genuinely novel — UI change, unexpected popup, ambiguous decision — it pauses, captures state, and pings a human queue. The human makes the call; the agent learns from the resolution.
Run nightly batch jobs. Trigger on inbound webhooks. Trigger on rows in your database. Trigger from Slack commands. The orchestration layer fits the operational shape of your business.
Multi-tenant isolation, encrypted credential storage, role-based access, audit logging, observability (Sentry + Better Stack), cost tracking on every model call, multi-LLM provider routing where applicable.
If we're going to charge you for a computer-use build, you deserve to know what you're getting and what the failure modes look like. Three things we're explicit about during scoping.
When the legacy vendor pushes an interface change (rare for the systems most clients use, but it happens), the agent's task definitions may need updates. Maintenance includes monitoring for these and reacting fast — but it's real ongoing work, not zero.
For high-stakes work (clinical decisions, financial transactions, legal filings), this is decision-support automation with structured review. The agent does the work; a human reviews the output. Full autonomy is reserved for lower-stakes, deterministic flows where the failure cost is low.
For HIPAA, financial-services, or other regulated contexts: BAAs in place, audit logging non-negotiable, data residency controlled. We document the posture for your records during scoping. If your regulators won't accept screenshot-based audit trails, we'll tell you on the discovery call.
We map the work currently being done in the legacy system — which tasks, how many hours per week, what's deterministic vs. judgment-driven. By end of call we know whether Computer-Use is the right answer or whether a different service (custom admin tool, document automation) fits better.
Stakeholder interviews, screen recordings of the target tasks, environment audit (which version, which OS, which login flow), success criteria defined per task, fixed-quote SOW. We rank tasks by automation feasibility — some are easy wins, some are not worth automating.
Weeks 1–2: VM environment, credential management, observability. Weeks 3–5: first task built and tested with thorough error handling. Weeks 6–8: additional tasks, orchestration layer, human-takeover queue. Weeks 9–10: hardening, edge cases, monitoring. Weeks 11–12: pre-launch dry-run with your team in parallel to existing manual process.
Two weeks of parallel run (agent + human in parallel) before full cutover. 30-day stability watch. Maintenance retainer covers VM hosting, monitoring, GUI-change adaptation, observability, cost tracking, LLM API baseline. Higher-touch retainer than other services because the maintenance work is real.
Traditional RPA records exact pixel-coordinate scripts that break the moment anything changes. Computer-use agents use vision models — they "see" the screen the way a human does, and they can adapt to small UI changes without breaking. Maintenance burden is meaningfully lower; capability ceiling is meaningfully higher (the agent can handle modal dialogs, error states, and judgment calls that RPA scripts can't).
Depends entirely on the task. Deterministic tasks against stable UIs run at high reliability — well above 99% across thousands of executions. Tasks with judgment or unstable UIs have lower reliability and require more human-takeover scaffolding. We measure this during paid scoping with a small pilot before quoting the full build.
Detected via the eval suite. Adaptation happens in maintenance — sometimes the agent adapts on its own (vision-model resilience), sometimes we update the task definition. Mean-time-to-recover is the metric we track, not "uptime in absolute terms."
Yes, for builds scoped that way. BAAs with all vendors that touch PHI. Encrypted credential storage. Audit logging on every action. Screenshot logs treated as PHI when applicable, with retention and access controls. The compliance posture is documented during paid scoping. If your specific regulators reject computer-use as an architecture, we'll tell you on the discovery call.
Three layers of defense. (1) Eval suite catches obvious failures pre-deploy. (2) Per-task validation (e.g., "after this task, confirm the saved record matches expected") catches in-production errors. (3) Audit log + screenshot trail let you reconstruct exactly what happened. For high-stakes tasks, decision-support framing keeps a human in the loop on every output.
Because the maintenance work is genuinely higher. GUI changes, OS updates, infrastructure issues, eval drift — all real ongoing work. We won't pretend otherwise to win the deal. The math still works: even at higher retainer cost, you're saving 10–20 hours of staff time per week.
For workflows that span legacy GUIs and modern APIs. The workflow orchestrator handles the modern systems; the computer-use agent handles the legacy ones. Same orchestration, two execution modes.
Discuss →For regulated practices where the legacy system is the system of record. The audit-prep agent reads from it (often via the computer-use agent) to maintain continuous compliance state.
Discuss →Sometimes the better answer is "replace the spreadsheet that's been working around the legacy system." Custom Admin Tool is the alternative when the right move is to leapfrog the legacy system, not automate clicking through it.
Discuss →A 30-minute discovery call confirms whether Computer-Use is feasible for your specific software, what the task list would look like, and whether a different service fits better.