Honest answers.

On disk, AES-256-GCM encrypted with a versioned key envelope (key_id rotation supported). The encryption key is derived from a server-side secret loaded only at boot and never logged. Plaintext leaves memory the moment it's saved.

The local LLM only sees your command (e.g. "refund order 1234 on Shopify") to pick a connector and parse arguments. It never sees vendor responses or other connectors' data. Routing runs on Steve's hardware — no third-party LLM API.

Sensitive actions (refunds, posts, DNS edits, deletions) land in an approval queue with the parsed intent shown in plain English. Read it, approve or reject. Reads execute immediately because they don't change state.

Those are workflow builders — you wire triggers to actions on a canvas, then maintain the wires. Commerce Claw is the opposite: type one sentence, the AI picks the connector. No canvas, no Zaps to maintain, no per-task pricing.

Free during early access. Eventual model is a flat workspace fee — no per-task metering, no credit math. Pricing changes will give 30 days notice; your audit log + connector registrations are exportable on the way out.

Not yet packaged for distribution. Architecture is one Node process + JSON store + local Ollama, so it lifts cleanly. Email info@agentabrams.com for the Docker recipe.

Most modern SaaS speak OAuth2 + REST and slot in cleanly — new connectors typically ship in 1-3 days. Anything that exposes a webhook can be wired the same week.

Encryption-at-rest, atomic writes with fsync, audit trail, sensitive-action approval queue, CSP+HSTS, rate limits — yes, by mainstream-SaaS standards. Early-access label is about connector matrix maturity, not security floor.