Kernax
Pricing → For Teams → CLI → Security → Docs → Blog →
Request Beta Access →
Docs / Configuration / Autonomy & Guardrails

Autonomy & Guardrails

Fine-tune agent privileges with kernel-level command interception, Hard-Lock destructive command blocking, and approval timeouts.

Granular Per-Tool Permissions

Autonomous agents should never operate with unchecked system access. Kernax enforces kernel-level command interception and granular privilege delegation directly inside your local environment:

Tool Capability Default Mode Configurable Behavior
File Reads Automatic (Always Allow) Limited to workspace directory; system root access is strictly denied
File Writes & Edits Interactive Review Can be toggled to Auto-Apply for trusted workspaces
Terminal Commands Confirmation Required Intercepted by Hard-Lock Guard before shell subprocess spawns
MCP Tool Invocations Per-Server Approval Can be whitelisted per MCP server in .kernax/mcp.json

Hard-Lock Destructive Command Guard

Hard-Lock Invariant: Kernax includes a hard-coded security heuristic running locally in Rust/C++. If an agent attempts to execute a destructive terminal pattern, the command is blocked instantly at the kernel layer, irrespective of autonomy settings.

Commands blocked by Hard-Lock include:

Approval Timeouts & Auto-Proceed

When running background Mission Control pipelines, you may configure how long the IDE waits for interactive confirmation before pausing or auto-proceeding:

settings.json
{
  // Seconds to wait for developer approval on sensitive actions
  "kernax.guardrails.approvalTimeoutSec": 45,

  // Action on timeout: "pause" (recommended) or "cancel"
  "kernax.guardrails.onTimeout": "pause",

  // Require explicit confirmation for external network curl requests
  "kernax.guardrails.confirmNetworkRequests": true
}

Unattended Agent Safety

If you dispatch a 30-minute background refactoring pipeline and step away from your workstation:

Was this page helpful?
Esc