I have five Claude Code sessions running right now. One is refactoring an auth module, one is stuck waiting on a test I forgot to write, one finished twenty minutes ago and I haven't noticed, and I genuinely cannot tell you the state of the other two without tabbing through five terminal windows.
That's the actual problem sir-claude is trying to solve, and it's the reason the solution ended up looking like a medieval kingdom instead of a dashboard.
The problem: parallel agents, serial attention
Claude Code is great at running several projects at once — spin up a terminal per repo, kick off a session, move on. The failure mode isn't the agent, it's you. Terminal tabs all look the same. A title bar doesn't tell you "this one needs a decision" versus "this one is still happily chugging through a migration." So you end up doing the thing humans are bad at: periodically re-checking N identical-looking windows to guess which ones matter right now.
Why a spatial metaphor actually helps here
A status dashboard with five rows of green/yellow/red text solves the same problem on paper, but it doesn't solve it in practice — rows of text all compete for the same kind of attention, which means you still have to read every row to know what's going on.
Spatial memory doesn't work that way. If "the auth project" is a specific building in a specific corner of a 3D scene, you don't need to read anything to know something changed there — you see a building get bigger, catch fire, or light up at the edge of your vision. That's the bet behind sir-claude: render each of your Claude Code project folders as a building in a small Warcraft II–style medieval kingdom, and let activity, status, and size do the reporting instead of text.
- Size = how active the project is — more recent commits/tool calls, bigger building.
- Color = status — healthy, waiting on you, stuck, done.
- Day/night cycle = ambient sense of how long things have been running, so a session that's been going all "day" reads differently than one that just started.
It's a goofy premise on the surface, but it's the same instinct behind why pilots use dials instead of a column of numbers — ambient, peripheral information beats information you have to actively go read.
How it's built
Under the hood it's a lot less magical than "medieval kingdom" makes it sound:
- A three.js scene renders one building per project folder. Each building is a self-contained mesh that reacts to a small state object (
activity,status,lastEvent) rather than owning any logic itself — the kingdom is a rendering layer, not where decisions happen. - A status → color/size mapping watches each project's Claude Code session and translates recent activity into the visual state above. This is the part that actually matters; the 3D is just the output.
- Clicking a building drops you into a per-project terminal plus a readable, parsed view of the Claude session history for that project — so the game layer is a map, not a replacement for the real interface. You zoom out to see the kingdom, zoom in to actually work.
- A "Spellbook" sits alongside each project for saved commands and scheduled jobs — the stuff you re-run often enough that typing it out each time is friction.
- "Conjure" is a natural-language-to-command layer: describe what you want in plain English, it proposes the actual command, you confirm before it runs. Keeps you in control without needing to remember exact flags.
- Offline voice control via
whisper.cppfor speech-to-text and Kokoro for TTS — no cloud dependency, no account, works the same on a plane as it does on your desk. - Telegram remote control, for checking on or nudging a session from your phone when you're away from the kingdom entirely.
None of this needs a server. No signup, no cloud dependency to run it — it's a local tool that happens to render its status screen as a game.
What's next
Two things I'm actively building toward:
- Proving Grounds — an LLM-vs-LLM arena inside the same kingdom, for comparing how different models/agents handle the same task instead of just trusting vibes.
- Deeper voice control — right now it's functional; the goal is for it to be the default way you check in on a running session, not a novelty feature.
Try it
npx sir-claude No install step, no account, no cloud dependency — it just runs against whatever Claude Code projects you already have locally.
The project itself is closed-source for now. If you want access to the source on GitHub, the easiest way in is to DM @sirclaude.dev on Instagram, connect via GitHub, or email ben.croughs@telenet.be — happy to talk through it.

