Skip to content
Why AGINDOWSee it. Understand it. Act.

Keep API changes from becoming broken workflows.

Designed to catch what changed, repair only what drifted, and show the evidence behind each decision.

YOUR AGENTAGINDOWVENDOR API
Pre-alpha / current reality

Local Rust engine and Node/Python SDKs. Both SDKs are unpublished. Detection is off until you configure it.

No production traffic. HTTP distributions are not deployed. Local tests establish local behavior, not production guarantees.

Read the full status table ↗
01 / The problemTwo sides of every call

The API changed.
Your agent didn’t.

A renamed field can stop a workflow. A changed unit can quietly write the wrong value. Valid JSON is only part of the contract.

Example language
Request drift

Synthetic example · field rename

// Agent sends
const request = {
  source: "card_42",
  amount: 2000,
};
# Agent sends
request = {
    "source": "card_42",
    "amount": 2000,
}
// Vendor now expects
const expectedRequest = {
  payment_method: "card_42",
  amount: 2000,
};
# Vendor now expects
expected_request = {
    "payment_method": "card_42",
    "amount": 2000,
}

A selective repair changes only the drifted field, when evidence and policy allow. The amount stays untouched.

byte_identical: false

A repair changes the payload. Its audit artifact must say so.

Response drift

Synthetic example · unit change

// Previously: total in cents
const previousResponse = { total: 2000 };

// Now: total in dollars
const response = { total: 20 };
# Previously: total in cents
previous_response = {"total": 2000}

# Now: total in dollars
response = {"total": 20}

The shape still fits. The meaning changed. Explicit statistical windows and invariants can supply evidence; a single response cannot establish a unit change.

Reads fail open. In repair mode, detected semantic write drift without a verified repair fails closed.

Illustrations only. These disclosures do not call an API, run detection or execute a repair.

02 / Build trustObserve → repair → strict

Earn the right to change a write.

Start with visibility. Enable repairs deliberately. Tighten policy as evidence earns your trust.

01

observe

Inspect and record. Never mutate or block the call.

02

repair

Apply only a verified, selective patch. Semantic writes without a safe repair are blocked.

03

strict

Refuse unrepairable write drift. Reads still fail open.

Original stays original. The no-repair path preserves the payload; repairs explicitly record changed bytes.

Engine faults pass through. Internal failures bypass inspection. Policy refusals are a separate decision.

Investigation stays outside the call. No synchronous database, LLM or other-cell dependency in the guard.

Read the policy and trust contract ↗
03 / Run it your wayLocal today

Start where your agent runs.

In process / local

Node & Python

Embedded Rust/WASM, local adapters and explicit guard configuration. SDK packages remain unpublished.

Node SDK ↗Python SDK ↗
HTTP / local

Sidecar

Loopback or Unix socket guard endpoint with explicit fail-open clients. Not deployed or production qualified.

Sidecar boundary ↗
Remote / not deployed

Hosted endpoint

A local Worker build exists. Hosted delivery and production qualification remain future work.

Build & limits ↗
04 / Show the workEvidence over promises

Read the proof.
And its limits.

Repair safety evaluationSynthetic internal cases; not production or detector calibration.Local performance measurementsScoped native/WASM fixtures; production latency remains unvalidated.Audit artifacts & verificationWhat changed, why, and what was checked. Acceptance alone does not prove write integrity.

Follow the implementation.

Explore the local SDKs and the decisions behind the runtime.

Open the project ↗

Dashboard preview: hosted login unavailable. Sign in ↗