H/M::AGENT SCOPELOCAL / LEAST PRIVILEGE / NO EXECUTION

FREE BROWSER TOOL / MCP PERMISSION PREFLIGHT

See the authority behind an AI agent.

Inspect which tools can read, write, delete, reach the network, execute code, move money or change identity before an agent receives them.

RUNS IN THIS BROWSER MODEL CALLS: 0 TOOLS EXECUTED: 0 RULESET::--
01

LOCAL CATALOG INPUT

Paste a tool list or load JSON.

TOOLS/LIST, TOOL ARRAY OR FUNCTION TOOLS / 200K
0 / 200,000

PERMISSION PREFLIGHT / WAITING FOR JSON

Tool metadata is a clue, not a control.

Agent Scope reads definitions locally, applies conservative MCP defaults and exposes where hard enforcement is still needed.

01
Parse tools and schemas
02
Map authority and combinations
03
Export a least-privilege review packet

HINTS INFORM / CONTROLS ENFORCE

Put every permission behind a harder boundary than the model.

MCP annotations improve review, but authorization, isolation, destination controls and human approval must live outside model reasoning.

  1. 01

    Minimize the catalog

    Expose only the tools required for the current task and remove abandoned capabilities.

  2. 02

    Scope identity

    Bind short-lived credentials to one resource, operation and task instead of sharing broad tokens.

  3. 03

    Isolate execution

    Restrict filesystem roots, commands and network destinations independently of tool descriptions.

  4. 04

    Confirm consequences

    Ask immediately before deletion, payment, publication, external messaging or privilege changes.

EXPLICIT LIMITS

A clean manifest does not prove safe behavior.

  • Annotations come from the server and may be incomplete, wrong or malicious.
  • Static JSON cannot prove authorization, sandboxing, token validation or runtime behavior.
  • Keyword analysis can miss novel or misleading tool descriptions and schemas.
  • The exported policy is a review artifact, not an enforcement engine or certification.