Description
OpenAdapt is an open-source project for turning a recorded human demonstration of a GUI task into a program that runs again on its own, and that only reports success, VERIFIED, when an independent check confirms the task's real-world effect actually took place, not just that the on-screen steps completed. The idea is that a person, or an agent they're watching, performs a task once; OpenAdapt compiles that demonstration into a repeatable program, a named human approves it, and a calling agent, such as a computer-use AI agent, can then invoke it going forward.
Read More
The compiled program and runtime live in a separate companion repository, openadapt-flow; this repository provides the installer and CLI (pip install openadapt, then openadapt flow tutorial). Running the tutorial records a task against a synthetic practice-management app called MockMed, compiles it, certifies it against a policy, and confirms the result through a read-only API separate from the screen that did the writing, in a few seconds and without calling a generative model at runtime. A companion command, openadapt quickstart --break-it, deliberately shows the tool refusing to report success when the underlying write is rejected, even though the on-screen banner shows a success message.
OpenAdapt distinguishes several possible outcomes for a run beyond simple success or failure, including VERIFIED, HALTED_BEFORE_EFFECT, RECONCILIATION_REQUIRED, and REJECTED_POLICY, each meaning something specific about what the system actually knows happened. It supports browser, native desktop (Windows, macOS, Linux), remote display (RDP) and Citrix/VDI surfaces, with different evidence available on each. Deployment options range from fully local or self-hosted use, to a customer-controlled model for sensitive data and private networks, to managed execution on OpenAdapt-run infrastructure.
The launcher, compiler, runtime, desktop app, substrate adapters and basic qualification tools are MIT-licensed and free to use; OpenAdapt Cloud is a separate commercial multi-tenant control plane for fleet governance, billing and enterprise integrations, and OpenAdapt Execute is a private partner service for running one qualified transaction inside an application a partner can't otherwise integrate with. The GitHub repository has around 1,700 stars and 260 forks as of this writing, with commits pushed within the last day, making it an actively maintained project rather than an abandoned one.
- CategoryWorkflow Automation
- Pricing ModelOpen Source
- DeploymentRun anywhere
- PlatformsWindows, macOS, Linux
OpenAdapt's Core Features
Compiles a recorded GUI demonstration into a repeatable, non-generative program
VERIFIED outcome requires an independent, out-of-band check of the real effect, not just an on-screen success message
Distinct outcome states (VERIFIED, HALTED_BEFORE_EFFECT, RECONCILIATION_REQUIRED, REJECTED_POLICY, and more) instead of a simple pass/fail
Supports browser, native desktop (Windows/macOS/Linux), RDP and Citrix/VDI surfaces
Local tutorial runs with no account, API key or model calls required
Optional PII/PHI scrubbing and platform-specific capture modules (Windows UI Automation, macOS Accessibility, Linux AT-SPI)
Workflow composition: chaining multiple compiled bundles with parameter handoffs between them
Local/self-hosted, customer-controlled, or managed-execution deployment models
MIT-licensed core, with a separate commercial Cloud control plane and a private Execute partner service
Getting Started with OpenAdapt
Install: Run pip install --upgrade openadapt (Python 3.10-3.12).
Check your environment: Run openadapt doctor --backend web to confirm dependencies like Chromium are ready.
Run the tutorial: Run openadapt flow tutorial to record, compile and certify a task against the synthetic MockMed app.
Author your own workflow: Use openadapt flow record, optionally with --url for a browser target, to capture a real task.
Compile and replay it: Run openadapt flow compile and openadapt flow replay to turn the recording into a tested bundle.
Serve it to an agent: Run openadapt-agent serve --allow-run so a calling agent can invoke the verified workflow.
OpenAdapt's Use Cases
- Verified back-office automation
- Computer-use agent hand-off
- Cross-surface workflows
- Regulated write operations
