Show it any repeated GUI task, once. It compiles into a governed, deterministic workflow.

OpenAdapt is a demonstration compiler for repeated GUI work wherever it lives — in the browser, in native desktop apps, or inside Citrix and other virtual desktops. Demonstrate a task once and OpenAdapt compiles it into a governed, deterministic, locally executable program that replays exactly, with zero model calls on a healthy run. When interfaces drift, OpenAdapt re-resolves from retained evidence or proposes a governed repair — and halts instead of guessing when verification fails.

OpenAdapt is for repeated work trapped behind browser, desktop, and virtual-desktop interfaces: too visual or variable for brittle selectors, but too consequential to hand to a free-form computer-use agent on every run.

Install OpenAdapt · Read the docs · See current limits · Visit openadapt.ai

pip install openadapt
openadapt flow demo-record --out rec
openadapt flow compile rec --out bundle --name mockmed-triage
openadapt flow certify bundle --policy permissive
openadapt flow replay bundle --run-dir run

This runs the bundled MockMed example and writes a human-readable REPORT.md. Use the five-minute guide to add lint, policy certification, drift, repair, and deployment.

Published comparisons of compiled replay against a computer-use agent, with the same external success check on both arms:

• Live third-party EMR (OpenEMR public demo, fake patients, 18-step add-patient-note workflow): compiled replay went 20/20 at 39.2s p50 with zero model calls; the agent went 10/10 at 70.4s p50 at about $0.55 of model charge per run. Small sample on a shared, daily-resetting demo — not CI-reproducible. Methodology and caveats.
• CI-reproducible control (bundled MockMed task): both arms passed every run (100/100 compiled, 20/20 agent), so the result is cost and latency, not success rate — 4.9s p50 with zero model calls versus 37.5s p50 for the agent. Methodology and caveats.

Zero model calls on a healthy run means no model-API charge on that run; it excludes authoring, review, maintenance, and infrastructure, and it is not a production-reliability or clinical-safety claim. Read the limits before extrapolating either result.

OpenAdapt is the Beta launcher and unified CLI; openadapt-flow remains the engine. Browser record, compile, replay, reporting, and bounded deterministic repair are the reference path. Desktop and remote-display backends retain their separate Experimental or Research labels.

Runnable is not the same as certified safe. Review identity coverage, postconditions, system-of-record effects, policy, and the published limits for each consequential workflow.

Model training, retrieval, grounding, and general computer-use work remain Research: openadapt-ml, openadapt-retrieval, and openadapt-grounding. They are not required for healthy deterministic replay.

OmniMCP, SoM, and PydanticPrompt are Labs, not product dependencies. Historical, Superseded, Deprecated, Archived, and Internal repositories are classified in the public lifecycle registry rather than presented as the product.

Product-engine changes belong in openadapt-flow. Packaging and launcher changes belong in OpenAdapt. Use each repository's issues for scoped work, or visit openadapt.ai for deployment inquiries.

Unless a repository says otherwise, OpenAdapt.AI code is MIT licensed.