OpenAdaptAI/OpenAdapt

OpenAdapt – Verified UI Automation for Clinical and Enterprise Workflows

What it is

• An open‑source Python‑based engine that records a human’s interaction with a desktop, web, RDP, or Citrix application, compiles that demonstration into a repeatable program, and runs it without any generative‑AI calls in the control loop.
• The key safety feature is a record‑and‑verify step: after the program types data into a target system (e.g., an EMR), OpenAdapt reads the record back through a separate read‑only channel (API, report, or second login) and checks that the entered values match what was intended. If the check fails, the run halts and a human is prompted to resolve the discrepancy.

Why it matters

• Many healthcare and enterprise teams still rely on manual data entry into legacy EMRs, payer portals, or remote‑desktop apps. Errors discovered later can cost time and money. OpenAdapt automates the entry and guarantees the result before it is accepted, turning a risky “click‑and‑hope” automation into a verifiable transaction.
• The system is designed to work with the exact UI your staff already uses—no custom integrations, no SDKs, and no changes to the underlying application.

Core capabilities

Capability	What it does
Demo recording	A person (or a local agent watched by a person) records a workflow on the target UI. OpenAdapt captures clicks, keystrokes, OCR, accessibility data, etc.
Compilation	The raw recording is compiled into a portable bundle that can be replayed on the same surface (browser, native desktop, RDP, Citrix). No generative‑model API is needed during execution.
Verification	After the automation writes data, OpenAdapt reads the record back via an independent source‑of‑truth (API, read‑only login, report) and confirms the effect. Outcomes are classified as VERIFIED, HALTED_BEFORE_EFFECT, RECONCILIATION_REQUIRED, etc.
Human‑in‑the‑loop fallback	If verification fails, the run stops before any change is committed and a single signed question is sent to a human operator for clarification.
Multi‑surface support	Works on web (DOM, accessibility, OCR), native Windows/macOS/Linux (UI Automation, AT‑SPI), RDP, and Citrix/VDI—all driven from a customer‑controlled runner.
Policy & contract enforcement	Workflows are certified against policies (e.g., clinical‑write). Effect contracts must be approved before a run can proceed to the next step.
Privacy‑first data handling	Recordings and screenshots stay on the runner’s machine unless explicitly shared. Optional openadapt[privacy] extras scrub PII/PHI.

How you can start

1. Install – python -m pip install --upgrade openadapt (Python 3.10‑3.12). The package pulls in Playwright for browser automation and optional extras for Windows/macOS/Linux UI capture.
2. Run the quick‑start demo – openadapt quickstart --break-it runs a synthetic clinic task twice, showing a successful run and a halted run where the backend rejects the save.
3. Record your own workflow –

   openadapt flow record --out rec               # local app
   openadapt flow record --url https://myapp   # web app
   openadapt flow compile rec --out bundle --name my‑workflow
   openadapt flow replay bundle --url https://myapp --run-dir run
   

4. Certify & deploy – Use openadapt flow certify with a policy (e.g., clinical-write) to generate an effect contract, then run the certified bundle in production. The runtime will emit a receipt showing whether the transaction was VERIFIED or required human reconciliation.

Enterprise options

• Hosted service – OpenAdapt offers a managed cloud runner for approved browsers and non‑sensitive workflows. Pricing and service details are on the OpenAdapt website.
• Customer‑controlled execution – For sensitive data (native apps, RDP, Citrix), you can run the engine on your own infrastructure; only metadata travels to the cloud control plane.
• Agent integration – The companion openadapt-agent package exposes compiled bundles over the MCP protocol, allowing external “computer‑use agents” to invoke verified workflows.

Evidence & reliability

• Public OpenEMR field test (July 2026) showed 19/20 runs passed OCR verification with zero model‑API calls; the 20th halted safely.
• Fault testing on a local test record system demonstrated that OpenAdapt reliably halts when records are missing, partial, duplicated, or written to the wrong patient.

Where to learn more

• Full docs: https://docs.openadapt.ai
• Live demo (synthetic runs): https://app.openadapt.ai/demo
• Community Discord: https://discord.gg/yF527cQbDG
• Source code for the compiler/runtime lives in the companion repos openadapt-flow and openadapt-agent.

---

Bottom line: OpenAdapt lets you automate repetitive data‑entry tasks in legacy clinical or enterprise systems while guaranteeing that the entered data actually made it into the system correctly. It does this with a transparent, auditable run‑report, human‑fallback handling, and a clear policy‑driven certification model, all under an MIT‑licensed, self‑hostable codebase.