Start with local evaluation
AGINDOW is a pre-alpha runtime guard for API calls made by AI agents. Published Node and Python packages run an embedded statistical guest. Configure comparable evidence windows to detect drift, then start in observe mode to preserve the original call. Installation, a successful call and a finding each establish different things.
Capability matrix
This matrix is shared with the product evaluation page. Available packages and locally tested behavior are distinguished from services you cannot use yet. A reference page describes an interface; it does not make that endpoint a public service.
| Capability | Evaluate today | Scope | Configuration | Limitation | Evidence |
|---|---|---|---|---|---|
| Local drift detection | Explicit statistical detection for magnitude (Unit), enumeration, completeness, type, charset / format. Public /try collects actual responses from an API you select and compares status checks, your supplied schema and the same packaged guest. No finding is forced; /docs/quickstart remains the SDK install path. | Embedded WASM guest in published Node/Python SDKs. Native library seams exist separately and are not the default guest or every SDK adapter. | Select an explicit detector with scoped baseline/live windows, await readiness, then wrap calls. Mode (observe/repair/strict) does not configure a detector. | The statistical guest is default; configure windows to detect. Without a selection it emits no findings. Detection against production traffic has not been calibrated. | Playground |
| Node / Python SDKs | Evaluate the published pre-alpha packages agindow@0.1.0-alpha.2 (npm) and agindow==0.1.0a1 (PyPI). Adapters, CLI report and parity fixtures exercise the shared WASM bytes. | Declared package metadata: Node >=22; Python >=3.11. Tested at Node v22.21.1 and Python 3.12 (relay CI); untested upper bounds are not claimed. | Registry install only. Contributors changing the engine still build the shared WASM guest. No hosted key required for embedded use. | Publication is not protection. Registry install does not enable detection, repair or hosted delivery. Self-serve Workloads registration requires an enabled host-wide Test/read policy and backend, with isolated profiles from verified organization membership; defaults remain unconfigured and local fixtures do not qualify hosted onboarding. | SDK docs |
| Repair proposals / verified repair | The Node/Python SDKs and native call path can apply selective patches for reviewed mechanical renames only after independent verification. Evaluate the configured mappings against your own payloads before enabling application. | Node/Python opt-in reviewed top-level field renames using the shared Rust repair library; ordinary inspection is non-mutating. | Explicit selectiveRepair/selective_repair rules, sufficient confidence, independent verification and repair/strict mode. Adapter install or mode alone is insufficient. | Proposed ≠ applied. Schema acceptance and confidence alone cannot establish intent. Ambiguous semantic mappings are not safe migrations. Selective repair is off by default. | Modes and repair limits |
| Production traffic use | Local policy matrix, availability bypass, audit composition and write-protection seams are implemented and testable in-repo. | Local library and Worker builds only. Sidecar and hosted HTTP distributions exist as explicit scripts; neither is deployed or validated against live production traffic. | Your deployment must provide the detector, verification callbacks and controls for turning protection on. Nothing turns itself on for production traffic. | AGINDOW does not currently operate a hosted write-protection service, and calibrated detection for production traffic remains unavailable. The homepage capability matrix records the remaining work without dates or unsupported promises. | Capability matrix |
| Tests / benchmarks | Reproducible local commands: cargo/vitest statistical and gated suites, internal repair-safety eval (42 synthetic cases), and local native/WASM performance fixtures. | Synthetic and local evidence only. Nothing has run against production traffic. | Follow published docs for install/quickstart baselines. Set hardware/runtime context when interpreting latency fixtures. | Internal repair-safety results are not detector accuracy. Latency numbers in docs are targets or local fixture means — not measured production latency. | Quickstart |
Supported versions and changes
The version dropdown lists Latest and retained earlier npm/PyPI releases. Select an earlier version to read its saved documentation and package pins. Switching keeps the same topic when that release has it; otherwise it opens that release’s documentation home. Search, navigation and Copy page follow the selected release. The SDK index lists the reviewed paired npm/PyPI release. Examples run against matching local distribution artifacts before publication. That validates those artifacts, not a fresh registry download on every supported platform. Follow the pinned version when reproducing a sample. Read the SDK changelog for release changes, the previous published pair and upgrade verification.
Frequently asked questions
Do I need an AGINDOW account? Embedded evaluation needs no separate AGINDOW credential. Does wrapping a call start learning? No: you supply explicit scoped windows; the SDK does not learn production baselines automatically. Does repair mode fix the finding? Selective repair is off by default; enable it with reviewed rename rules, an independent verifier and repair/strict mode. Can I send calls to agindow.com? No managed guard service is currently available. Keep vendor credentials in your application’s existing client.