Trust Center

The Full Stack. Every Vendor. Every Control.

This page exists for the security teams, compliance officers, lenders, and legal counsel who need the long version. Owner-operator finance does not get the same documentation that institutional finance gets, by default. So we wrote it ourselves. Every control we use, every vendor we depend on, the data flow end to end. Ask us anything we did not cover.

Platform + Practice

What We Use From Anthropic, and What We Build Ourselves

Manage runs on Anthropic's generally-available Claude products with a Zero Data Retention arrangement on the inference layer. The vertical practice on top is MSIQ-built. Honest about which is which.

What we use from Anthropic

Generally-available Claude products, configured for your engagement:

  • Claude Messages API with Zero Data Retention. The inference layer for every skill output. ZDR means Anthropic does not retain your prompts or outputs at rest after the API response is returned. The arrangement is per-organization and provisioned via Anthropic's sales team.
  • Claude Agent SDK. The Python and TypeScript libraries Anthropic ships for building agent loops. Our skills are built on this SDK directly. ZDR is preserved because the SDK uses the Messages API underneath.
  • Claude Cowork. Anthropic's collaborative writing surface ($20/mo per seat). Where Scott and your team draft monthly close narratives, board decks, and lender packages with Claude as a real-time collaborator.
  • Claude for Excel. The Microsoft Excel add-in ($20/mo per seat). Where your CEO and growth team edit driver-based forecasts and scenario models live, with Claude tracing references across tabs and diagnosing model structure.
  • Claude on AWS Bedrock with PrivateLink (enterprise add-on). Available for clients who require network-isolated deployment and have an existing AWS posture. Not the default; scoped per engagement.

Cowork and Claude for Excel are not currently ZDR-eligible by Anthropic's policy. We use them transparently for collaborative work; high-sensitivity flows run on the ZDR-eligible Messages API.

What we build ourselves

The vertical practice and the security layer on top:

  • Credential broker. MSIQ-built local secrets store with OS-keychain-backed encryption. Issues short-lived per-call tokens. AI context never holds real refresh tokens, so a prompt-injection attack cannot exfiltrate credentials. ZDR-compatible by design.
  • Pre-prompt PII redaction. Customer names, account numbers, EINs, and other identifiers are tokenized before any data reaches the LLM. The model reasons about tokens; only authorized humans see the un-tokenized output.
  • Per-client read scoping and structured audit log. A decorator on every skill enforces single-client reads. Every skill run writes a JSONL audit entry to your encrypted client folder. You can read it directly.
  • 49 vertical-specific skills. Marketplace settlement reconciliation, cohort economics with long-window variants, marketing halo attribution, multi-location inventory, 3PL invoice forensics, cash-forecast drift detection.
  • Four vertical packs. Ecommerce, wine, health and wellness, elective medicine. Different stacks, different cohort definitions, different compliance calendars.
  • Per-client CEO dashboard. Four tiers, phone-first, with the disclosure tier discipline (real / benchmark / fixture) we ship on every artifact.
  • The pre-publish gate. Four-checker integrity chain (canonical facts, scope, voice, factual verification) on every artifact that leaves the engagement. Plus Scott signs every monthly close.
The Six Controls

Three From Anthropic. Three From Us.

Every embedded finance engagement should be able to name its security controls specifically. Here are ours.

Zero Data Retention on the Inference Layer

The Claude Messages API is ZDR-eligible by Anthropic's published policy. Under a ZDR arrangement, your prompts and outputs are not stored at rest after the API response is returned. The arrangement is per-organization and provisioned through Anthropic's sales team. We apply for ZDR before any client integration touches live data.

No Training on Your Data

Anthropic does not train on commercial API customer data. This is contractual under the API terms of service and applies independently of ZDR.

Optional PrivateLink via AWS Bedrock

For enterprise clients requiring network-isolated inference, we deploy Claude on AWS Bedrock with PrivateLink. Your queries route through your own AWS VPC; financial data does not transit the public internet. Added as a scoped per-engagement option, not the default. See pricing and architecture spec on request.

MSIQ Credential Broker

Tokens for QuickBooks, Shopify, Amazon, bank feeds, and ad platforms live in an MSIQ-built local secrets store backed by OS-level keychain encryption. Skills request short-lived per-call tokens via placeholder substitution. AI context never holds real refresh tokens, so a prompt-injection attack cannot exfiltrate credentials.

MSIQ Pre-Prompt Redaction + Client Scope

Customer names, account numbers, EINs, and other identifiers are tokenized before any LLM call. A decorator on every skill enforces single-client read scoping so cross-client access is structurally impossible.

MSIQ Audit Log + Pre-Publish Gate

Every skill run writes a structured JSONL audit entry to your encrypted client folder, accessible to your team directly. Every artifact bound for outside our engagement (CPA, lender, board) passes through a four-checker gate: canonical facts, scope, brand voice, factual verification.

Claude Cowork and Claude for Excel, which we use for collaborative document work, are not currently ZDR-eligible by Anthropic's published policy. We disclose this transparently and route high-sensitivity flows through the ZDR-eligible Messages API path.

Regulated Engagements

HIPAA Readiness and Data Residency

HIPAA. For elective medicine clients handling protected health information, we run engagements under a signed Business Associate Agreement on a HIPAA-enabled Anthropic organization. HIPAA readiness is provisioned at the organization level via Anthropic's sales team. Note: HIPAA-ready API access is not available on AWS Bedrock; PHI engagements run on the first-party Claude API with the BAA.

Data residency. Standard engagements process data in US-based Anthropic infrastructure. Bedrock-deployed clients choose their AWS region directly.

Subprocessor list. Our full subprocessor list is maintained at our Privacy page and updated whenever we add or change a vendor.

For The Buyer Who Probes

A Note On Claude for Financial Services

Anthropic released "Claude for Financial Services" on May 5, 2026. It is a product for institutional finance firms: Citadel, Carlyle, BNY, Travelers, Walleye Capital, Hg. It ships ten agent templates designed for investment banking workflows (pitchbook construction, KYC screening, model building, earnings review) and a set of data-partner integrations (FactSet, S&P Capital IQ, MSCI, PitchBook).

Manage does not use that product. The institutional-FS agent templates are not designed for owner-operator finance and the data-partner integrations are not relevant for the buyer we serve. We use the underlying generally-available Claude products (the Messages API, the Agent SDK, Cowork, Excel) directly. If a security team asks whether MSIQ is "on the same platform as Citadel," the answer is honest: same Anthropic platform underneath, very different product surface on top.

Anything We Did Not Cover?

If your security team, compliance officer, or legal counsel has a specific question, write to security@mainstreetiq.com. We answer in writing, on the record, with the specifics you need to take the engagement to your reviewers.

Book an Intro Call → Send a Message