HiveKey
Scope · what each agent can do

Give each agent exactly what its job needs. Nothing more.

Scope is least-privilege for agents. Define the capabilities as a reusable role, apply it across the fleet, and keep everything you didn't grant completely invisible — the agent can't call or even see a tool it wasn't given.

The model

Capabilities are a privilege, not a default.

Most agent stacks hand over a raw API key — every tool, all the time. Scope inverts that: an agent gets a deliberate, auditable surface and nothing else.

Least privilege by default

New agents start with zero capabilities. You add surface deliberately, one tool at a time.

Invisible, not just denied

Ungranted capabilities are hidden from the agent's tool manifest — closing the prompt-injection door before it opens.

Roles, not snowflakes

Govern by job function. One change to a role updates every agent that wears it — no per-agent drift.

Role → permission matrix

One grid, the whole org's agent permissions.

Rows are roles, columns are capabilities. A glance tells you who can move money, who can touch production, and who can only read.

access-matrix · production

5 roles · 6 capabilities
Role mail_sendpayments_payvault_getcrm_readcrm_deletedeploy
Support · L2
Billing
Ops · deploy
Read-only
Admin

Illustrative matrix. granted · not granted (invisible to the agent)

Role editor

Edit a role; every agent wearing it updates.

No code, no key rotation. Toggle a capability and the change propagates across the fleet on the next action.

roles / Support · L2
draft

mail_send

Send mail to approved domains

crm_read

Read CRM records (no PII export)

vault_get

Read secrets from the vault

payments_pay

Initiate payments

crm_delete

Delete CRM records

Applies to 12 agents

Publish role

What the agent sees

With vault_get off, that tool never appears in the agent's manifest. A prompt-injected request to "reveal the API key" has no tool to call.

// agent.tools (resolved)

[
  "mail_send",
  "crm_read"
]

vault_get · payments_pay · crm_delete
not exposed

How it works

From zero to a governed agent in four steps.

1

Define a role

List the exact capabilities the job needs — mail_send, crm_read — and nothing else. Roles are version-controlled and reviewable.

2

Assign it to agents

Apply the role to one agent or a hundred. The agent inherits exactly that surface; changes to the role propagate instantly.

3

Everything else is invisible

Ungranted tools don't appear in the agent's tool list at all. It can't call — or even discover — what it wasn't given.

4

Adjust without redeploy

Tighten or widen a role in the console. No code change, no key rotation, no agent restart.

Put every agent your company runs under one policy.

Watch HiveKey scope, guard, and block a live action on your own agents — 30 minutes, no slides, no commitment.