Skip to content

OpenAdapt

OpenAdapt turns a repetitive task in an application without a usable API into a program that enters the data and then checks that it saved. A person or an agent shows the task in the real application. OpenAdapt compiles one or more demonstrations into a workflow program you can read, tests it in your environment, and runs it with new inputs. A healthy run makes no AI model calls. When a check fails, the run stops and asks a person instead of guessing.

It drives web apps, Windows, macOS, and Linux apps, and apps reached over RDP or Citrix: a CRM, a billing system, an insurance portal, or an electronic health record. The tutorials use a synthetic clinic app as the example. The engine is open source under the MIT license.

For developers

  • Run it in one command

    uvx --python 3.12 openadapt quickstart --break-it

    Or install it on Python 3.10 to 3.12 and run the engine tutorial:

    python -m pip install --upgrade openadapt
openadapt flow tutorial

    No account or API key. Quickstart tutorial

  • Give a workflow to your agent

    claude mcp add openadapt -- uvx --python 3.12 openadapt-agent serve

    The MCP server gives an agent each compiled workflow as one tool call. This starts a sandbox that needs no account. Connect an agent

Jump to: Command line · How the compiler works · MCP server · Run outcomes · Your first workflow · Security & procurement

How a save is checked

On the acting surface, the app screen OpenAdapt types into, a run checks that the right record is on screen, then enters and saves. It then reads the record back. A screen read-back uses the app's UI in a separate session and is not independent. An independent read uses the system of record through its API, database, or file export, never through the UI. If the read matches, the run ends VERIFIED with a run report and a receipt; OpenAdapt Cloud signs a Seal only after an independent read. If it doesn't match or can't tell, the run ends RECONCILIATION_REQUIRED and a person checks the record. If the right-record check fails, the run ends HALTED_BEFORE_EFFECT before saving and asks a person.

The screen that did the work can show "Saved" when nothing was stored, so it never confirms a save on its own. How a save is checked explains which read-backs count for which result.

From demonstration to production

flowchart LR
    S["Show the task<br/>(a person or an agent,<br/>one or more times)"] --> C["Compile<br/>a workflow program<br/>you can read"]
    C --> P["Certify<br/>policy and lint checks"]
    P --> Q["Qualify<br/>normal and fault cases<br/>in the target environment"]
    Q --> A["Admit<br/>this exact version"]
    A --> R["Run<br/>right-record check, action,<br/>record check at each save"]
    R --> V(["VERIFIED<br/>report, receipt, Seal"])
    R --> H(["Stopped<br/>HALTED_BEFORE_EFFECT"])
    R --> U(["Not sure it saved<br/>RECONCILIATION_REQUIRED"])
    H -.->|"a person or an agent<br/>shows the fix"| C
    U -.-> K["A person checks<br/>the record"]

How OpenAdapt works walks through each stage. Terms are defined once, in the glossary. Which releases are admitted for production is on Release integrity.

When it fits

OpenAdapt fits a task when most of these are true:

  • The team does it the same way many times a week.
  • What to enter is already decided, by a person, a form, or an AI agent.
  • The application has no practical API for the write.
  • A wrong entry is costly.
  • There is a second way to read the saved record, such as an API, a database, a report, or a read-only sign-in.

If the application already has an API for the task, call the API. The limits page lists what OpenAdapt doesn't do.

What the packages are

On Python 3.10 to 3.12, pip install openadapt installs the openadapt command, the engine (openadapt-flow), and the MCP server (openadapt-agent). The packages map lists every package and repository and which ones you need.