openadapt-flow

OpenAdapt is built for repeated workflows behind browser, legacy desktop, and remote-display interfaces. It compiles demonstrated GUI workflows into deterministic, locally executable programs. Each target application and environment is qualified separately. Healthy runs make no model calls. When interfaces drift, OpenAdapt re-resolves from retained evidence or proposes a governed repair and halts when verification fails.

Real screenshots from the two runs in docs/showcase/. Left: the UI the demo was recorded on. Right: a theme it had never seen — each step re-resolves through OCR or geometry, and each fix is written back to the script as a reviewable diff. Zero model calls on either side.

Verified execution. It halts instead of guessing, and qualification reports measure silent incorrect success, over-halt, effect confirmation, latency, and model calls. Read the technical limits and validation method, including five adversarial rounds against the wrong-target check.

Try it

pip install openadapt-flow

openadapt-flow demo-record --out rec                     # record a demonstration
openadapt-flow compile rec --out bundle --name my-task   # compile it
openadapt-flow lint bundle                               # expected: finds demo gaps
openadapt-flow certify bundle --policy permissive        # smoke-policy pass
openadapt-flow certify bundle --policy clinical-write    # expected: strict refusal
openadapt-flow replay bundle                             # replay: local, $0
openadapt-flow replay bundle --drift theme \
  --save-healed-to healed                                # deterministic repair

On the first command that needs a browser, openadapt-flow downloads the Chromium build Playwright needs (a one-time ~150MB fetch) — no separate playwright install chromium step. Prefer the fast, isolated installs uvx openadapt-flow … or uv tool install openadapt-flow. In air-gapped or CI environments that pre-provision the browser, set OPENADAPT_FLOW_NO_AUTO_INSTALL=1 to disable the auto-download.

The bundled MockMed tutorial is intentionally runnable but not certified for clinical writes. lint exits nonzero because its irreversible final click is unarmed, and clinical-write refuses additional identity, system-effect, and idempotency gaps. That is the safety boundary working, not a setup failure. The permissive policy is only a smoke gate. Replay serves MockMed and writes report.json, an illustrated REPORT.md, and reviewable repair patches under heals/.

The nightly clean-machine test runs this complete install-to-uninstall journey on Linux, macOS, and Windows. See the capability and qualification matrix for the accepted scope of each substrate.

Record your own app

record --url opens a headed browser on YOUR app and watches what you do — real clicks, typing, key presses and scrolls — writing the same recording format compile consumes. Perform the workflow, then press Ctrl-C (or close the window) to finish:

openadapt-flow record --url https://your.app --out rec   # do the task, Ctrl-C
openadapt-flow compile rec --out bundle --name my-task
openadapt-flow replay bundle --url https://your.app       # replay it

Pass --url to replay to run against your own app; recorded parameter values are the defaults and --param overrides them.

Secrets never get recorded. A input[type=password] field (or any field named with --secret <name>) is a secret parameter: its value is never written to the recording, the events log, the compiled bundle, or the saved frames (its region is redacted). At replay it is injected from the environment and a missing one fails fast:

openadapt-flow record --url https://your.app --out rec --secret password
export OPENADAPT_FLOW_SECRET_PASSWORD='…'                 # supplied at replay
openadapt-flow replay bundle --url https://your.app

Compiled is not the same as certified safe. lint reports a bundle's coverage gaps (clicks that act with no identity check, steps that assert nothing, write steps left mis-classified) with a severity each; certify enforces a policy and exits nonzero — refusing the bundle before it deploys — when it fails. Risk is auto-classified at compile time (write-shaped clicks — save/submit/create/delete/... — become irreversible, which arms the low-confidence refusal), and two example policies ship: a permissive default and a strict clinical-write.yaml. See docs/LIMITS.md for what the heuristic does and does not catch.

How it works

Computer-use agents re-reason through your task with a large model on every run. That's the right shape for a task nobody has automated before, and the wrong one for the 500th referral this month. openadapt-flow compiles the demonstration instead.

Each compiled step carries a template crop, an OCR label, geometry landmarks, a structural locator, and postconditions derived from what the demo actually changed on screen. At replay time a resolution ladder tries them in order: a structural element match where the backend owns a DOM/UIA tree, then local template match, global template match, OCR, landmark geometry, then (optionally) a grounding model. Healthy scripts normally resolve on the first rung. Individual deterministic resolution steps complete in milliseconds; end-to-end workflow time depends on the target application. The healthy path makes no model calls and incurs no per-run model cost.

When bounded UI drift preserves enough evidence, a lower rung can find the same target and the fix lands in the bundle as a diff you can review. An optional model may propose a repair only when explicitly enabled; a human can teach a guarded correction after a halt. These are different modes, not a blanket promise of adaptation. When the screen stops matching expectations entirely, the run halts with a report instead of guessing, and steps tagged irreversible won't act on a low-confidence match at all.

The runtime is vision-first: it can operate a pure pixel surface (PNG in, clicks and keys out), but it is not limited to pixels. Where a backend owns a structured layer — a browser DOM, a native UI Automation / accessibility tree — the ladder's top rung re-finds the recorded target as an element and acts on it deterministically; the visual rungs are the fallback floor for pixel-only substrates (RDP, Citrix, canvas). On a desktop drift benchmark the structural rung resolved 21/21 targets where visual replay alone managed 6/21 (benchmark/structural_action/). Structure never bypasses the identity gate — it makes identity stronger, an exact element rather than a pixel guess. But the identity gate only covers armed steps, and today's bundles arm a subset of clicks (the live OpenEMR bundle armed 4-7 of 12) — an unarmed click has no identity check at all. The per-step coverage is auditable in workflow.json and reported in every run; see what it doesn't do yet.

It all sits behind a small Backend protocol shared by browser, native desktop, and remote-display drivers. The browser lifecycle runs on every CI build and has published third-party application evidence. Windows UIA passed 3/3 fixed WinForms trials with independently confirmed SQLite effects, plus 3/3 refusal for both stale and ambiguous targets. Native macOS passed 3/3 fixed TextEdit trials with exact file-byte effects and refused a two-window ambiguity without changing either file. Real-network Aardwolf RDP into Windows 11 passed 3/3 fixed remote-input trials with independent guest-tools file verification. These are accepted scoped qualifications; a customer application is qualified against its own controls, session/display policy, identity evidence, and effect oracle. Native Linux now uses the same governed runtime through an exact-window AT-SPI backend. Required CI drives a real GTK3 workflow through an isolated X11/session-D-Bus environment for three independently verified effects, plus three ambiguous-target and three stale-target refusals; customer applications retain their own qualification boundary. Citrix ICA/HDX uses the remote-display adapter contract and enters design-partner qualification in the customer's exact published application (docs/backends/RDP.md, docs/desktop/LINUX_NATIVE.md, docs/desktop/CITRIX_PIXEL.md).

Proof

Every CI run records a demonstration, compiles it, and checks:

Scenario	Outcome
Baseline replay ×3	all steps template rung, 0 heals, 0 model calls
Theme drift	succeeds; 8/8 anchors healed; healed bundle replays clean
Moved buttons	succeeds via global template search
Renamed buttons	succeeds via landmark geometry
Surprise modal	fails loudly, naming the violated postcondition
Non-recorded parameter	substituted and verified by OCR of the final screen

Artifacts: baseline run report · theme-drift run report.

Compiled workflows can also be emitted as Agent Skills or MCP servers (emit-skill / emit-mcp), so other agents can invoke them.

From trace to program

A single demonstration under-specifies intent, so openadapt-flow does not stop at replaying one. These capabilities layer onto the same $0, model-free runtime:

• A workflow program, not just a line of steps. Beyond the linear v0 bundle, the IR (openadapt_flow/ir.py) expresses a parameterized program: states and guarded transitions, loops over worklists, subflows, typed parameters, predicates, and exception paths (ProgramGraph / State / Transition / LoopSpec / Guard / Predicate / ParamSpec). The flat trajectory is the degenerate case, so the migration is backward-compatible. Design: docs/design/WORKFLOW_PROGRAM_IR.md.
• Multi-trace induction that refuses when it isn't sure. induce_program aligns several demonstrations of the same task to recover the shared parameters, loops, and branches — deterministic and model-free at its core. When a branch condition or a value stays underdetermined it quarantines the program (certified is False) instead of guessing, and disambiguate surfaces the ambiguity as concrete multiple-choice questions rather than inventing an answer.
• Effect verification against the system of record. The screen can lie: an optimistic UI, a duplicate submit, a partial save all read as success. A step may declare typed effects, and when a run is given an EffectVerifier the replayer checks the real record — REST (RestRecordVerifier), FHIR (FhirEffectVerifier), or a document hash (DocumentHashVerifier) — before and after the action, halting on a refuted or unverifiable write, still with zero model calls. A fault-model study found the screen-only oracle silently mishandles 5 of 7 transactional fault classes; all five halt through the real replayer once effects are declared (benchmark/fault_model/, docs/design/EFFECT_VERIFIER.md). Two honest preconditions bound this: the compiler does not yet infer effects from a demonstration — they are authored per deployment against the app's system of record — and a run with no verifier configured falls back to the screen oracle. The net exists only when both are supplied; without them the write is exactly as silent as before.
• An API actuator tier. Where the target app exposes a real API, driving its GUI to make the write is the wrong tool. A step carrying an ApiBinding, with an ApiActuator configured, performs the write by calling the API deterministically and confirms it with the same EffectVerifier — the api leaf of the capability ladder (API → DOM/UIA → geometry → OCR → template → VLM → human). It is an optimization whose safe fallback is always the GUI.
• Policy: lint and certify. lint reports a bundle's coverage gaps (unarmed clicks, vacuous postconditions, under-classified risk) with a severity each; certify enforces a policy and exits nonzero, refusing a bundle before it deploys. Runnable is not the same as certified safe. Certification is optional and opt-in — an uncertified bundle still runs — and a policy only defines what a bundle must satisfy, so the honest claim is that a certified workflow can be configured to halt on the conditions its policy names, not that any workflow always halts.
• Governed healing. Every fix under drift lands in the bundle as a reviewable diff, and a step classified irreversible will not act on a low-confidence match — structure and the identity gate govern the heal, they are not bypassed by it.
• Durable checkpoint / resume. A run checkpoints verified progress (openadapt_flow/runtime/durable/) so a halt becomes a durable pause the operator can approve and resume from the last verified state — not a restart, and explicitly not "hand the rest to a free-form agent."
• PHI-free identity. The wrong-patient identity check can run against a salted-hash, shape-preserving IdentityTemplate instead of a plaintext name / DOB / MRN band, so a compiled bundle need carry no readable PHI while still enforcing identity (openadapt_flow/runtime/identity_template.py).

Benchmark

The lead result is on a real third-party app: the official OpenEMR public demo (fake patients only, resets daily). We ran an 18-step add-patient-note workflow both ways — log in, find a patient, scroll a dense dashboard, add a note — with a distinct note value each run and the same OCR success check on both arms: 20 compiled replays against 10 runs of a claude-sonnet-5 computer-use agent. Compiled went 20/20 at 39.2s (p50) with zero model calls; the agent went 10/10 at 70.4s (p50), about $0.55 per run at list price ($5.52 total for the 10 runs, with prompt caching and hard cost caps enforced in the harness). It's a shared public demo that other users mutate and that resets daily — not CI-reproducible, and the sample is small. Correctness alone (no agent arm, 5/5 fresh browsers, zero model calls, closed-loop scrolling) is in docs/showcase-openemr/FINDINGS.md. Full numbers, methodology, and caveats: benchmark/openemr/BENCHMARK.md.

For a controlled, CI-reproducible comparison — the methodology anchor — we ran the bundled MockMed task both ways on 2026-07-08 with the same OCR success check: 100 compiled replays against 20 runs of the same agent. Both arms went 100 for 100 and 20 for 20, so on an app this simple the story isn't success rate. It's that a compiled replay finishes in 4.9s (p50; 5.1s p95) with zero model calls, while the agent takes 37.5s (p50; 43.4s p95). The measured agent sample cost about $0.27 per run at the model's then-current list price; repeat-run figures are projections and exclude authoring, maintenance, and infrastructure. Full numbers, methodology, and caveats: benchmark/BENCHMARK.md.

The stack also ships a pinned, containerized lending reference environment, benchmark/frappe_lending/ — pinned containers + lockfile, with independent REST, SQL, and exact table-delta verification of every write. In the model-free engineering matrix (compiled and direct-API arms, baseline plus cosmetic drift), it delivered 12/12 correct rows with zero silent wrong writes, zero over-halts, and $0 model cost. The paid agent arm and full per-cell trial counts are the next stage of the matrix.

Capability and qualification

The reference browser path runs record, compile, policy-check, deterministic replay, refusal, and report generation in CI. Windows UIA, native macOS, and RDP each have retained 3/3 accepted task evidence with independent effects or oracles. Citrix and each new third-party application are qualified with a design partner in the exact deployment environment. The workflow-program IR adds parameters, branches, loops, effect verification, and governed recovery on the same runtime. DESIGN.md has the module contracts; docs/design/WORKFLOW_PROGRAM_IR.md describes the program IR, and docs/L1_INTEGRATION.md covers feeding layered clinical-data platforms.

The integrated status of the engine, browser, desktop, remote-display, safety, GUI, hosted, and deployment surfaces is published in docs/PRODUCT_STATUS.md. Security reviewers should start with docs/ENTERPRISE_ARCHITECTURE.md, which maps screenshot/credential flows, cryptographic guarantees, hosted boundaries, and unmet controls.

Machine-checked claims. Product claims are enforced by CI. Every registered claim is tiered and mapped to the specific tests and benchmark artifacts that back it in claims.yaml. CI runs scripts/validate_claims.py, which fails the build whenever a claim's tier outranks its strongest evidence and regenerates docs/VERIFICATION.md — the claim-by-claim verification report — from the registry, so the adjectives in this README cannot quietly rot.

Privacy (PHI)

For regulated deployments, PHI scrubbing on the persist/log paths is provided by the optional privacy extra (Presidio-backed openadapt-privacy):

pip install 'openadapt-flow[privacy]' && python -m spacy download en_core_web_sm
export OPENADAPT_FLOW_SCRUB=on          # scrub REPORT.md + logs, fail closed

The shareable REPORT.md and console logs are scrubbed; the compiled bundle and report.json keep literal identifiers on purpose (identity check + audit trail) and are protected by a documented boundary. Identity crops sent to the on-prem VLM appliance are deliberately not scrubbed — the control there is on-prem-only + no-retention. Full map: docs/PRIVACY.md.

At rest, opt-in AES-256-GCM encryption (OPENADAPT_BUNDLE_KEY) seals workflow.json, template crops, and durable checkpoints. KMS integration and key rotation remain operator responsibilities, and full-disk encryption is still required. Treat every source bundle as PHI. Details: docs/phi_at_rest.md.

Hosted (cloud connectivity)

Hosted commands connect the locally executed compiler/runtime to the launched control plane at app.openadapt.ai: authentication, governed artifact ingest, and PHI-minimal break reporting. Mint an ingest token in the dashboard (<host>/dashboard/settings/ingest), then:

pip install 'openadapt-flow[privacy,hosted]'
openadapt-flow login --token oai_ingest_…
openadapt-flow sanitize ./my-recording --kind recording --out ./triage.sanitized
openadapt-flow review-sanitized ./triage.sanitized --original ./my-recording
# add missed redactions locally, then approve in the viewer or CLI:
openadapt-flow approve-sanitized ./triage.sanitized --original ./my-recording \
  --reviewer operator@example.com
openadapt-flow push ./triage.sanitized --kind recording

# Compile only from the approved sanitized recording, then validate locally.
openadapt-flow compile ./triage.sanitized --out ./triage.bundle --name triage
openadapt-flow lint ./triage.bundle --strict
openadapt-flow certify ./triage.bundle --policy permissive
openadapt-flow replay ./triage.bundle --url https://example.internal/login \
  --run-dir ./triage.run --param patient_id=example

# Privacy-review the executable bytes. A changed executable is refused.
openadapt-flow sanitize ./triage.bundle --kind bundle --out ./triage.bundle.sanitized
openadapt-flow review-sanitized ./triage.bundle.sanitized --original ./triage.bundle
openadapt-flow approve-sanitized ./triage.bundle.sanitized \
  --original ./triage.bundle --reviewer operator@example.com

# Bind exact artifacts and local evidence to a one-time Cloud challenge.
openadapt-flow validate-hosted \
  --recording ./triage.sanitized --bundle ./triage.bundle.sanitized \
  --run-dir ./triage.run --policy permissive --risk-class low \
  --environment staging-v1 --target-url https://example.internal/login \
  --out triage.validation.json
openadapt-flow push ./triage.bundle.sanitized --kind bundle \
  --validation-attestation triage.validation.json

# To activate this as a new version of an existing workflow, add:
#   --workflow-id 00000000-0000-0000-0000-000000000000
# To bind that replacement to the exact halted run it repairs, also add:
#   --resolves-run-id 00000000-0000-0000-0000-000000000000

openadapt-flow report-break runs/replay-… \          # PHI-free break diagnostic
    --workflow-id <id> --deployment-kind byoc         #   → POST /api/runs/ingest-report

• Token resolution (all outbound calls): --token → OPENADAPT_INGEST_TOKEN env → OS keychain → existing ~/.openadapt/config.toml token (migration read). Install the hosted extra for keychain storage. New plaintext mode-0600 storage is refused unless login --allow-plaintext-token makes the insecure fallback explicit.
• Sanitization never mutates the original. It inventories every file, applies type-specific text/image handlers to a copy, requires a stable second scrub pass, and writes per-file source/derivative hashes and coverage to .openadapt-sanitization.json.
• Review is local-only by default. review-sanitized binds to 127.0.0.1, loads no remote assets, presents original and derivative side by side, accepts additional literal/rectangle redactions, and invalidates prior approval after every change. Administrators may opt into policy approval only for fully covered, stable derivatives. Automatic hosted approval additionally requires a deployment-allowlisted HMAC signing key; an ingest token cannot self-assert that policy.
• Approval freezes exact bytes. It creates a deterministic immutable archive and binds reviewer, policy, timestamp, SHA-256, and byte size. push sends those exact bytes plus the openadapt.sanitization/v1 manifest; it never re-zips after approval.
• Destination trust is independent of deployment lane. OpenAdapt's managed origin is recognized explicitly. A customer-managed/BYOC endpoint requires HTTPS plus an exact-origin allowlist. Sanitized artifacts may upload from cloud, BYOC, regulated, or PHI-mode lanes; unknown destinations are refused.
• Current coverage is text and still images. Symlinks and database, video, audio, nested archive, encrypted, executable, or unknown files are refused, never copied through or reported as covered. See docs/SANITIZED_ARTIFACTS.md.
• Sanitizing a bundle can break execution. If a load-bearing target, typed value, identity crop, or postcondition changes, the manifest marks runtime semantics unvalidated and push --kind bundle refuses it. Parameterize PHI before compilation or execute the original inside its trusted boundary.
• Runtime validation is separate from privacy approval. It binds the exact approved recording and bundle, compiler configuration, parameter schema, strict lint, named certification, derived low/consequential risk class, successful report, and exact HTTPS target/host boundary to a short-lived, one-time tenant/token challenge. Cloud also requires exact deployment allowlist membership for certification policy, derived risk class, and a compiler version actually deployed by the runner. The HMAC proves token possession and envelope integrity; it is not independent observation, a compliance certification, or a safety SLA.
• Halt signaling is read from report.json (RunReport.halt / HaltObservation), never from a process exit code (replay/run return 0/1 only). report-break posts only a schema-minimal descriptor: hashes, status, resolver rung, and numeric metrics. Free text, screenshots, DOM, and field values never enter the automatic payload. A 422 boundary rejection retries the same minimal shape, then falls back to local-only.
• Opt-in post-run hook: set OPENADAPT_FLOW_HOSTED_WORKFLOW_ID (and optionally OPENADAPT_FLOW_DEPLOYMENT_KIND / OPENADAPT_FLOW_ORG_ID) and a halting replay/run emits the break automatically (best-effort; never changes the run's exit code). Off by default.

The sanitizer uses the optional privacy extra. Hosted transport uses httpx.

Development

git clone https://github.com/OpenAdaptAI/openadapt-flow && cd openadapt-flow
pip install -e '.[dev]'
playwright install chromium  # optional: else auto-downloads on first launch
pytest -q

Contributions are welcome — see CONTRIBUTING.md. A ready-made first contribution: pick a module off the mypy type-debt burn-down list ([[tool.mypy.overrides]] in pyproject.toml), tighten its annotations, and remove it from the list.

The demo GIF is generated from real run artifacts by scripts/make_demo_gif.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 remain within the declared MIT package boundary.