openadapt-flow

openadapt-flow records a task you do once in a web or desktop app and compiles it into a program that runs on your own machine. A healthy run makes no AI model calls. It's built for data entry that repeats every week and has to be right, such as claims, referrals, eligibility checks, and invoice posting.

Two checks guard the writes. Before a click on a record, the run compares the text around the target, such as the patient name in that row, with the recording, and stops if it differs. After a save, a production run reads the record back through a second channel, such as an API, a read-only SQL query, or a report export. It reports success only when the record matches what it entered.

It drives web apps through Playwright; Windows, macOS, and Linux apps through their accessibility APIs; and Citrix or remote desktops through screen pixels. Agents can call a compiled workflow as a tool over MCP.

Docs · Try it in your browser · Website · Discussions

See it run

pip install 'openadapt-flow[browser]'
openadapt-flow demo

The demo runs one recorded task twice against a fake clinic app on your computer. The first run uses an honest app. The second uses an app that shows "Encounter saved" and then drops the note. When both runs finish, the demo opens a page with the two results side by side. The first run can take a few minutes while Playwright downloads Chromium. With the OpenAdapt launcher installed, openadapt flow demo runs the same command.

Synthetic recording. A screenshot of the page that openadapt-flow demo wrote on 2026-10-09 against the bundled fake clinic app; regenerate it with scripts/make_demo_proof_image.py.

The two final screens are identical, pixel for pixel. Only the record check tells the runs apart. To watch each stage in the terminal instead, run openadapt-flow tutorial, then openadapt-flow tutorial --break-it; docs/TUTORIAL.md shows the output.

How it compares

The time and cost rows come from one matched comparison on MockMed, the bundled fake clinic app: 100 compiled runs and 20 agent runs, all scored by the same OCR check of the final screen (method). The agent was claude-sonnet-5 with the computer-use tool, driving the same browser through screenshots, and cost is model API usage at list price. On the public OpenEMR demo, compiled replay completed 19 of 20 runs (the 20th stopped before saving) with a 39.2 s median; the agent completed 10 of 10 with a 70.4 s median at about $0.55 per run (method). Both comparisons ran on 2026-07-08 with Flow 0.1.0. The Playwright column describes Playwright's documented behavior: we haven't timed a codegen script on the same task or run one through our UI-change fixtures. A side-by-side comparison with UiPath, Power Automate, and browser agents is at openadapt.ai/compare.

When the UI changes

Each recorded step keeps several kinds of evidence for its target: the DOM element or accessibility node where the surface has one, an image crop, the label text, and the target's position relative to nearby landmarks. At replay the run tries them in that order and uses the first match. A step that resolves through a fallback is logged as a proposed repair under the run's heals/ folder. --save-healed-to writes a repaired copy of the bundle for review; Flow never swaps it in by itself. When nothing matches, the run stops instead of clicking something plausible. If an irreversible step, such as a final submit, can only be found by position, the run refuses to click. On a step with an identity check, the target goes through that check before the click, however it was found.

These are the changes the tests cover, on MockMed or a synthetic test page. Except where a row says otherwise, the tests switch off the element locator, to test the fallbacks that pixel-only surfaces such as Citrix depend on.

Left: the UI the task was recorded on. Right: a dark theme the recording never saw. Neither run makes a model call. The run artifacts are in docs/showcase/, and the order of fallbacks is in docs/RESOLUTION_LADDER.md.

Try it on your app

Record the task once against a test copy of your web app, compile it, and replay it:

pip install 'openadapt-flow[browser]'
openadapt-flow record --backend web --url https://your-test-app.example --out rec
openadapt-flow compile rec --out bundle --name my-task
openadapt-flow lint bundle
openadapt-flow replay bundle --url https://your-test-app.example

Python 3.10 through 3.12 is supported. On Windows cmd.exe, use double quotes: pip install "openadapt-flow[browser]".

lint lists what the bundle doesn't cover yet, such as a write with no record check. It exits 1 only for an error; a fresh recording usually has a few warnings. replay is a rehearsal: it runs the steps but doesn't read the record back. It ends COMPLETED_UNVERIFIED (finished, not checked) and exits 0, so exit 0 from replay doesn't mean the save was confirmed.

To have OpenAdapt set up a workflow with your team, start at openadapt.ai.

Add the record check

A production run reports success only after it reads back every write it declared. Each workflow needs two things:

1. A description of what each write changes. openadapt-flow scaffold-verifier rec drafts one from the recording's write steps. The draft checks nothing until a person edits it.
2. A second way to read the record. A database connection isn't required. A read-only REST, FHIR, or GraphQL API works, and so does a read-only SQL query, a file, report, or email the app exports, or, when you have none of those, the app's own record page opened in a separate signed-in browser session (ui_readback, a weaker check). You configure it once in deploy.yaml, and nothing changes in the app. See docs/EFFECT_KIT.md.

Reading the screen the run just used doesn't count, because an app can show "Saved" and drop the write, as the demo shows. Without a second read, Flow can still rehearse the workflow, but a production run can't report a checked save: it stops, or it hands the result to a person.

openadapt-flow certify bundle --config deploy.yaml
openadapt-flow run bundle --profile standard --config deploy.yaml

certify checks the bundle against a policy. clinical-write is the strict built-in policy, and permissive is a smoke check. A new recording usually fails clinical-write until its contracts are complete; lint lists the gaps. run --profile standard checks the policy again, replays, and reads the record back.

Exit codes

For anything a script acts on, read transaction_outcome in report.json. Only VERIFIED is a checked save, and only HALTED_BEFORE_EFFECT, REJECTED_POLICY, CANCELED, and FAILED_PLATFORM prove nothing was written. RECONCILIATION_REQUIRED means a person checks the record before anything is retried.

Surfaces

Install the extras for the surface that records and replays the workflow:

docs/SURFACES.md has the record and replay commands for each surface, and docs/PRODUCT_STATUS.md lists the evidence for each one. A bundle is bound to the surface it was recorded on.

One bundle drives one surface: one browser tab, one desktop app, or one remote desktop session. For a task that crosses surfaces, record one bundle per surface and chain them with openadapt-flow compose. Each child starts only after the one before it ends with a checked save, unless you allow a specific stop with --allow-halt.

Call a workflow from an agent

pip install 'openadapt-flow[execute]'
openadapt-flow serve-execute --port 8787 --seed-mockmed

This starts a local HTTP server on loopback. It speaks MCP at POST /mcp, with tools to start a run (create_execution), poll it (get_execution), and fetch its signed result (get_execution_receipt). The same operations are available as plain HTTP under /v1/executions, and requests carry the bearer token the server generates on first start. The server runs only bundles registered in its data directory, each pinned by its digest, and runs them under the standard profile, so the result says whether the record check passed. --seed-mockmed registers two simulated MockMed workflows, an honest app and one that drops the save, which return results without opening a browser.

openadapt-flow emit-mcp bundle --out server.py writes a standalone MCP server for one bundle. It replays without the record check.

From Claude Code

examples/claude-code-demo records one form fill in a small expenses app backed by SQLite, compiles it, and serves it to Claude Code over MCP with openadapt-agent. Claude Code passes a description and an amount. OpenAdapt fills in the form with no model calls, then reads the expenses table itself. The tool returns success only when exactly one new row with those values is there. If the app shows "Expense saved." without writing the row, the tool returns halt.

examples/claude-code-demo/demo.sh   # install, start the app, record, compile
cd examples/claude-code-demo
claude --mcp-config mcp.json        # then ask: Add an 18.75 expense for parking

What runs where

Recording, compiling, replaying, and production runs all happen on your machine. A healthy run makes no model calls, and model-assisted target finding is off unless you turn it on in the CLI or deployment config. That isn't the same as no network: the app you drive and the record check still talk to their own services.

For regulated data, the privacy extra scrubs PHI from the shareable report and the console logs:

pip install 'openadapt-flow[privacy]' && python -m spacy download en_core_web_sm
export OPENADAPT_FLOW_SCRUB=on

The bundle and report.json keep literal identifiers, because the identity check and the audit trail need them (docs/PRIVACY.md). With OPENADAPT_BUNDLE_KEY set, openadapt-flow seal ./bundle --out ./bundle-sealed encrypts workflow.json and the image crops with AES-256-GCM. Other bundle files aren't encrypted, so treat every source bundle as PHI (docs/phi_at_rest.md).

OpenAdapt Cloud at app.openadapt.ai is an optional hosted control plane for browser workflows. Desktop and Citrix runs are self-hosted (docs/HOSTED.md).

Limits

• If the app has an API for the task, call the API.
• Each workflow takes setup: a recording, a description of its writes, and a second read path. For a task you'll run a few times, doing it by hand or pointing an agent at it costs less.
• A record check sees only what its read path covers. In a fault test with injected bad writes, a check that trusted the success message passed 54 of the 72 runs that left the record wrong. One API read let 9 of the 72 through, all writes to a table it didn't read (EFFECT_E2E.md).
• The identity check runs only on steps where the recording captured identifying context. The live OpenEMR bundle had it on 4 to 7 of 12 clicks.
• In our tests on the pixel path, any change in zoom, display scaling, or font size stopped the run at its first step.
• The Citrix path was tested on a stand-in with no DOM (3 healthy runs and 3 stops on drift), not a live Citrix ICA/HDX session. Each deployment has to be tested on its own Citrix setup.
• A stop isn't a rollback. Steps before the stop may already have changed the app.
• The comparison numbers above are from July 2026, on Flow 0.1.0, with 20 or fewer agent runs per task. They haven't been measured again on a later release.

The full list, capability by capability: docs/LIMITS.md.

More documentation

• Production setup: qualifying a bundle from its recording, policies, multi-step processes, signed results, and release status.
• docs/CAPABILITIES.md: workflow programs with branches and loops, for-each over a worklist, induce from several recordings, durable checkpoint and resume, and Agent Skill output.
• docs/BENCHMARKS.md: every benchmark, with method and caveats.
• docs/REWARD_WORKER.md: serve-reward scores a model's training episode by reading the record after it ends.
• docs.openadapt.ai: the full documentation set.

Development

git clone https://github.com/OpenAdaptAI/openadapt-flow && cd openadapt-flow
pip install -e '.[dev]'
python -m playwright install chromium  # optional browser pre-provisioning
pytest -q

On Linux, Flow checks the Chromium host libraries before an automatic browser download. A minimal host may need a one-time system-library install. Flow stops before the download and prints a command for the exact Python environment that runs it.

Contributions are welcome; see CONTRIBUTING.md. For a first contribution, pick a module off the mypy type-debt list ([[tool.mypy.overrides]] in pyproject.toml), tighten its annotations, and delete it from the list.

The demo GIF is generated from real run artifacts by scripts/make_demo_gif.py. The proof image at the top of this page comes from a real openadapt-flow demo run through scripts/make_demo_proof_image.py.

License

OpenAdapt-authored package code is licensed under the MIT License.

A Git checkout or GitHub-generated source archive also contains an isolated openIMIS reference environment with adapted configuration files under AGPL-3.0-only. The MIT license does not relicense those files. Their exact provenance, file-local scope, and complete upstream license are recorded in the repository-only third-party notice.

Published PyPI wheels and source distributions exclude the openIMIS benchmark surface and stay within the declared MIT package boundary.