Bernato Blog

Best Tools for Verifying AI-Generated Code Changes in 2026

Manual review, CI, static analysis, LLM review bots, and cryptographic provenance each catch a different kind of failure. None of them alone is enough. Here is the honest breakdown.

Jul 11, 2026

The Case for a Local-First Agent Fleet (and When Cloud Orchestration Wins Anyway)

Local-first wins on cost, privacy, and latency for most individual developers. Here is exactly where that stops being true, with a calculator to find your own crossover point.

Jul 6, 2026

Dial-Out WebSockets Explained: How Bernato Reaches Your Agents Without Opening a Port

No port forward, no UPnP gamble. The daemon dials out to a broker and holds the socket open, so your home machine never has to accept inbound connections.

Jun 25, 2026

Signed Provenance for AI Coding Agents: How to Prove a Diff Wasn't Hallucinated

"It compiles" proves nothing. Signed provenance hashes and signs every step from prompt to diff, so anyone downstream can verify the code came from the run it claims to.

Jun 14, 2026

Approval Fatigue Is Real: Designing Permission Prompts Agents and Humans Both Trust

If every agent action pops an approval, people start rubber-stamping. Here is what actually changes that: narrow scope, batching, and a prompt that says exactly what's being asked.

Jun 3, 2026

How Many Claude Code Agents Can One Machine Actually Run?

The real bottleneck usually isn't CPU. Use the calculator to estimate a safe concurrent-agent count for your own RAM and cores.

May 23, 2026

Devcontainers vs Bernato Sandboxes for Running Untrusted AI Code

Devcontainers isolate with a full container filesystem. Bernato uses native OS sandboxing with capability tokens. Here is when each one actually wins.

May 12, 2026

PM2 vs systemd vs Bernato for AI Agent Fleets

PM2 and systemd keep a process alive. Neither knows what a capability token is or gives you a live terminal you can approve from your phone. Here is the real comparison.

May 1, 2026

Kill Switches for AI Agents: Design Patterns That Actually Stop a Runaway Process

SIGTERM is a request, not a command. Here is how to design a kill switch with a grace period, a hard fallback, and a master switch that works when the dashboard doesn't.

Apr 20, 2026

Capability-Based Permissions: A Safer Model Than Full Disk Access, Yes or No

Binary permission prompts train people to click yes. Capability tokens let you grant an agent exactly what it needs, nothing more, and log every grant.

Apr 9, 2026

Is ngrok Safe for Your Home Server? An Honest Risk Assessment

ngrok is convenient but routes your traffic through a third-party relay. Here's the actual threat model, what ngrok sees, and when self-hosted alternatives are worth the switch.

Mar 30, 2026

Run Claude Code on Every Machine You Own

A laptop, a GPU rig, a homelab box, an old Mac mini under the TV. Here's how to run Claude Code on all of them and supervise everything from one dashboard, or your phone.

Mar 26, 2026

Sandboxing Claude Code on macOS: Stop Yolo Mode Without Losing Power

--dangerously-skip-permissions is fine until it isn't. Here's how to wrap Claude Code in macOS Seatbelt with capability tokens.

Jan 17, 2026

Keep a Claude Code Session Alive After Closing the Terminal

Claude Code's Remote Control documentation says 'your terminal must remain open.' Here's how to make that not be true, using a daemon that owns the agent, not your shell.

Dec 11, 2025

Git Worktrees + Bernato: Parallel Claude Code Sessions With Full Supervision

Git worktrees solve 'agents fighting over node_modules.' Bernato solves 'where is each agent running.' Combine them and you have N independent CC sessions, all visible from one screen.

Nov 4, 2025

The Audit Log Every AI Agent Needs

If something weird happens, can you say what the agent did? A local append-only activity log is the difference between investigation and guesswork.

Oct 16, 2025

Tailscale vs Cloudflare Tunnel vs Bore: What I Use For My Daemon's Tunnel

Best ngrok alternative that keeps data on your own hardware. Honest review of three reverse-tunneling stacks: when each is right.

Oct 6, 2025

When Kubernetes Is Overkill: A Process Orchestrator for Homelab

Best process manager for homelab servers in 2026. K3s alternative for homelab enthusiasts who don't need pod scheduling.

Aug 4, 2025

What Is Zero-Trust Process Orchestration?

A short, opinionated definition. A control plane for spawning, supervising, and killing processes across machines you own, where every action is authenticated and every decision is logged.

Jul 26, 2025

Best Process Manager for Homelab in 2026: PM2, Supervisor, systemd, Bernato Compared

Six process managers ranked by what they're actually good at, with the decision matrix that says which one for your homelab.

Jul 8, 2025

LAN-Direct Routing for Local Dashboards

When you're on the same Wi-Fi as your daemon, the dashboard talks to 127.0.0.1 directly and falls back to a tunnel only when needed.

Jun 26, 2025

Restart-on-Crash for Long-Running AI Agents

AI agents crash. Models 500. Network blips. Here's a small ceiling-and-backoff design that stops infinite respawns without giving up on transient errors.

Apr 18, 2025

Pair a Machine in 30 Seconds Without SSH Keys (Capability Tokens)

SSH key pairing is two-factor. Capability tokens collapse the manual step into a one-time token issued by the tray.

Apr 16, 2025