Skip to main content
Operational guardrails

Prove to your security team that AI can't break your ERP.

Every request is checked before it reaches your ERP: who is calling, which service they own, whether their credentials resolve, and whether the change must pause for human review. The AI never gets a direct line to your system.

Request guided access

Write gate

No surprise mutation

human review

pending confirmation

{
  "tool": "update_purchase_order",
  "operation": "update",
  "target": "PurchaseOrder(4500008123)",
  "expiresIn": "10m",
  "requires": ["review", "confirm"]
}
argument hash checkedexecutor app-only

Trust comes from the request path.

Each request passes through identity, ownership, credential resolution, operation classification, confirmation, and audit logging.

01

Authenticate

Validate caller identity and token path.

02

Authorize

Check organization and service ownership.

03

Resolve

Load credentials only when MCP connection requests need them.

04

Classify

Identify read, write, function, workflow, or app-only execution.

05

Confirm

Pause sensitive writes for human review.

06

Audit

Record actor, decision, payload state, and result.

Multiple auth modes share one MCP connection boundary.

The MCP router extracts organization, service, session, and token context, then checks access before delegating to the OData bridge.

OAuth 2.1 bearer token

Standards-compatible MCP authentication.

Basic auth

Maps platform login to a platform user context.

Supabase JWT

Reuses browser or SSO session identity.

Static MCP token

Automation and CLI access with revocation.

Credential storage pattern

Store secrets in Supabase Vault
Save only secret references on instances
Resolve through an Edge Function on demand
Avoid process-level credential caching
Avoid logging secret values
Rotate by updating Vault

Credential shapes

Basic username/password
Cookie string
OAuth bearer token
OAuth client credentials
No credential required

What cannot run without approval.

High-impact operations are legible, attributable, and stoppable before they change records.

See confirmation in workflows
Gate 1

Create ERP records marked as confirmation-required

Gate 2

Update exposed fields on protected entities

Gate 3

Delete records through generated mutation tools

Gate 4

Execute a workflow step configured to pause for review

Defense in depth

Platform-wide assurance.

Mutating function imports and actions are explicitly classified as safe or confirmation-required before they can run.

Custom UI execution flows through the same authorization and audit path as direct tool calls.

Credentials flow exclusively through the vault resolver — direct references never reach the process layer.

Every AI request is intercepted by the DataTether backend (identity confirmed, operation scope checked, OData call assembled) before your ERP receives it. The AI never has a direct line to your system.

How visual views (MCP UI) stay governed

Audit-ready records

Confirmation IDs, actors, tools, payload state, decisions, expiry, and execution status give operators a record of what happened and why.

Harden your first connection

See every control in action: scoped access, approval gates, and audit trails from day one.

Governed by default. Auditable from the start.

Bring one OData service. We will show you the exact exposure boundary, credential handling, confirmation gates, and audit path before a single AI action touches your enterprise data.

Months of dev time eliminated

Auto-generated tools replace weeks of custom MCP server development.

No infrastructure to run

Fully managed. Nothing to install, host, or maintain.

Instant dashboards and UI

Live tables, KPIs, and review screens — no frontend developer needed.

One setup, every AI client

ChatGPT, Claude, Manus, custom agents — all from the same connection.

A good first step is small: one service, one user outcome, one visible control boundary.

Request guided access