Automate the UI-only work your APIs can't reach.
OpenAdapt compiles demonstrations into governed workflows across browser, native desktop, RDP, and Citrix. Healthy runs execute deterministically and make no model calls. Consequential actions are identity-gated, results are checked against the workflow's evidence contract, and uncertainty halts for review instead of being reported as success.
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 · Watch the live demo · Read the docs · Visit openadapt.ai
The flagship OpenAdapt repository
has more than 1.6k stars, reflecting interest in the overall OpenAdapt project;
the canonical current engine is
openadapt-flow.
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 runThis 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 head-to-head comparisons, each graded by an external success check that is independent of every arm:
- 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.
- Independent effect verification (fault-injection study, 90 runs per arm, end to end through the real replayer into an on-disk SQLite system of record, graded by a direct read-only database connection that bypasses the service): a screen-only "success banner" oracle silently accepted 75.0% of the wrong effects that actually occurred (54 of 90 runs). Adding one out-of-band verifier that reads the system of record cut that to 12.5% (9 of 90). A complete read path over every mutable surface reaches 0 of 90, but the realistic deployment number is the middle rung — one out-of-band oracle — not the 0%. All nine residual misses are a single named class: a collateral write to a surface the oracle does not read. Every run terminates in an explicit transaction outcome (VERIFIED, HALTED_BEFORE_EFFECT, RECONCILIATION_REQUIRED, and others), so uncertain delivery is surfaced for reconciliation, never reported as success. 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.
| Surface | Lifecycle | Start here |
|---|---|---|
| Open-source product | Beta | OpenAdapt installs the unified CLI; openadapt-flow is the canonical compiler and governed runtime. |
| Desktop cockpit | Beta | openadapt-desktop provides local recording, qualification, execution, evidence review, and governed repair. Native installers are published from its releases page. |
| Cloud control plane | Beta | app.openadapt.ai provides managed browser execution plus workflow, evidence, attended-operation, usage, and billing surfaces for customer-controlled execution. Its implementation repository is private. |
| Native capture | Experimental | openadapt-capture is the canonical native screen, input, timing, and window-scoped recording component used by Desktop and Flow. |
| Agent bridge | Beta | openadapt-agent exposes compiled Flow workflows as local MCP tools and Agent Skills without becoming a second runtime. |
| Documentation | Beta | docs.openadapt.ai is the canonical journey-led site; openadapt-ops is its Internal publishing source. |
| Evidence and evaluation | Research | openadapt-evals contains evaluation infrastructure. Runnable references live in openadapt-flow/docs/showcase, with methods and evidence under benchmark. |
OpenAdapt is one Beta product across browser, Windows, macOS, Linux, RDP, and Citrix/VDI. Those execution surfaces share one qualification contract while retaining environment-specific bindings and evidence. Qualification remains specific to the workflow, application, environment, identity contract, and effect verifier rather than being inferred from a platform checkmark.
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.
