AI governance control plane

Govern what AI may do. Keep the record of what it did.

Capability Index is a governed gateway between your AI assistants and your real actions. Every capability carries a permission tier — run it, ask first, or hold for a human — enforced the same whether the caller is Claude, Copilot, n8n, or a script.

How it works

One gateway between assistants and your systems

Assistants don't touch your systems directly. They speak to Capability Index, which for every call answers three questions in order — who is calling, may they use this, and what ceremony does it need — then runs it (or doesn't) and writes it to the register.

Claude Copilot n8n API token over MCP / bearer Capability Index GATEWAY 1 · Identity — who's calling 2 · Access — may they use it? 3 · Permission — auto · confirm · verify Your systemsAPIs · n8n · webhooks Audit registerwho did what

One request, three checks, then run + record — identical on every surface.

The assistant never sees your credentials. A capability tells the model what it does — the endpoint, token and secrets stay server-side and are resolved only when the action actually runs. Nothing sensitive is in the model's context or logs.

One governance, many surfaces

Author a capability once. Govern it everywhere.

The same capability, the same permission tier, the same audit trail — whoever calls it. The surface is just transport; the governance lives in the gateway, so you write and review it in one place.

Claude · MCP Copilot Studio · MCP n8n · API token Your app · REST Scripts · API token

The gate

Three tiers, from “just run it” to “hold for a human”

The permission tier is declared per capability and enforced at the gateway — not left to the assistant. A mutating action can never fire on an unconfirmed call.

Request Tier?declared auto Runs ▶ confirm preview → human yes Runs ▶ verify staged for operator released → runs

The assistant can narrate the tier, but only Capability Index decides whether the action runs.

Auto
Runs

Low-risk, reversible actions run immediately and are logged.

Confirm
Asks

Shows exactly what will happen and waits for a human yes.

Verify
Holds

Staged for an operator to review and release on the console.

A worked example

“Pay a €400 refund” — the safe path

1

The assistant asks

An agent (or an n8n flow) calls the pay_refund capability with an amount and a customer.

2

Capability Index holds it

The capability is marked verify. Nothing runs — it's staged, and the caller gets back “pending approval,” not a result.

3

A human releases it

An operator reviews the exact action on the console and releases it. Only then does it run — and the whole chain is in the register.

What you get

01

A permission on every action

auto, confirm, or verify — declared per capability and enforced at the gateway, so a destructive action can't fire without a human.

02

One control plane, any surface

The same governance over MCP to any assistant, plus scoped API tokens for machines like n8n — identity, access, and tiers hold everywhere.

03

Evidence by default

An audit register of who did what, with what outcome — the artifact you hand an auditor, assembled from data the gateway already holds.