Randy Northrup

Sysadmin · Network & A/V · Web Dev. I build AI Agents, MCP Servers & Dev Tools.

[Randy Northrup @RandyNorthrup]

I'm Randy: sysadmin, network and A/V tech, and web dev in Ventura, CA. I build AI agents, MCP servers and dev tools in the open.

A quick tour of what I've shipped:

[Randy Northrup @RandyNorthrup]

11 Tailscale and Headscale flaws went public this week without coordinated disclosure.

If you run either one: did you change anything (Tailscale SSH off, tighter ACLs), or just update and move on?

[Randy Northrup @RandyNorthrup]

@_MoustafaM Do you route everything through Amnezia, or only the connections from networks that block plain WireGuard?

[Randy Northrup @RandyNorthrup]

@joelhooks Do the agents post to that network in plain text a human can skim, or structured messages? Curious whether you read the feed yourself or only the summaries.

[Randy Northrup @RandyNorthrup]

@deirradecoded @claudeai Did it map gaps per domain, or all the way down to individual objectives? Curious how it handled the blueprint items the cert guide barely covers.

[Randy Northrup @RandyNorthrup]

@juanhenaoparra What happens when two task branches touch the same files? Does the gateway flag the conflict, or is that left to whoever merges?

[Randy Northrup @RandyNorthrup]

@Haackas0 Does the game tell players spare parts are being lost while the pile's full, or do they find that out on their own?

[Randy Northrup @RandyNorthrup]

The home lab corner 🛠️

Standing desk, ASUS monitor + a portable screen, Stream Deck, Tascam on a mic stand. Underneath: an Avatar PD705 pad, kick + hi-hat pedals and a pair of JBLs.

Lighting is rough. What's on your desk that you'd never give up?

[Randy Northrup @RandyNorthrup]

@TheHackersNews Until a fix lands, the sysadmin move is shrinking blast radius: turn Tailscale SSH off on boxes that don't need it, fall back to plain sshd with keys there, and audit your ACLs so only admin devices can reach port 22 at all.

[Randy Northrup @RandyNorthrup]

@je_rwl A countdown is a fair way to do it. Does it tell you which mode is coming, or just that something is?

[Randy Northrup @RandyNorthrup]

@DannyLimanseta Simple and it works. Does the .bat rebuild before it launches, or do you still kick off the build from Claude first?

[Randy Northrup @RandyNorthrup]

@RoyXanderide Would you want a nudge when you land on a near-synonym, or would that spoil the fun of finding the one that works?

[Randy Northrup @RandyNorthrup]

Been building a word game with Claude's help. This is its real title screen: each prop on the table is a nod to one of the game's worlds, and plays a little hint when you click it.

Meet Baffleword.

[Randy Northrup @RandyNorthrup]

@dani_avila7 Good pick for repro work. For the cases where you need your real logged-in session, Claude helped me build chrome-control-mcp (MIT, on my GitHub): it drives your everyday Chrome through an extension, and each editor gets its own tab so a few agents can share one browser.

[Randy Northrup @RandyNorthrup]

@business Opposite use case here: Claude helped me build an OBS plugin that turns Ray-Ban Metas into a live POV camera, straight over the glasses' own Wi-Fi, no phone in the loop. Camera-free would kill that one for me.

[Randy Northrup @RandyNorthrup]

@neil_xbt Was the eval() one caught by a scanner rule in the skill, or by the agent actually tracing where the input goes? My quality_retrofit skill (MIT, on my GitHub) uses red drills for the same worry: break the code on purpose and make sure the gates fail.

[Randy Northrup @RandyNorthrup]

@witcheer On (3), does it read the standard MCP progress notifications, so any server that already sends them just lights up? Or does a server need something Hermes-specific to show the percentage?

[Randy Northrup @RandyNorthrup]

@GenAI_is_real That fight is why Claude helped me build chrome-control-mcp (MIT, on my GitHub). Clicks go in at the browser level so your mouse stays yours, and each agent gets its own tab without stealing focus. Would you share your Chrome like that, or give the agent its own desktop?

[Randy Northrup @RandyNorthrup]

@SadCreatorTalks Worse if the agent pulls issues through the API instead of the rendered page, since the raw markdown keeps HTML comments. Treating issue text as data and never as instructions has to be the default now.

[Randy Northrup @RandyNorthrup]

@afscott Does Claude drive Resolve through its scripting API, or hand it an EDL/XML for the finish? That handoff seems like the part that decides how much cleanup is left.