The workshop puts a few long-running AI agents ("desks") in one room, on the same work, so you direct a team instead of relaying between chat tabs. This takes you from nothing to a running desk in a couple of minutes.
- Workshop — a shared workspace: a repo, or a local folder. Everything in it shares one bench.
- Desk — one long-running agent with its own memory and history, opened inside a workshop. You can put several on the same problem.
- You — the operator. You set direction and make the calls the room can't settle on its own; the board tells you what actually needs you.
Reach for it when a problem wants more than one angle, over time — a security review, a gnarly migration, a research question. For a quick one-off, just use the CLI; you don't need a room for a single shot.
Quickest (Windows): run Start.bat from the repo root — one thing to
run. It builds, launches the tray (a cairn icon by the clock), starts the server
hidden, and opens the dashboard in your browser:
Start.bat
Prefer a live console instead? Run the web app directly:
dotnet run --project src/WorkshopRoom
…or start the tray launcher on its own (click the cairn icon to open, quit its menu to stop everything):
dotnet run --project src/WorkshopRoom.Tray
Open the URL it prints (a localhost port). The board reads your live Copilot
sessions straight from ~/.copilot/session-state and refreshes on its own every
few seconds. It opens in simple view — just your desks, grouped by workshop,
with tokens and signals.
On first run the setup is open and the board is empty:
Two ways to start (step ①):
- Create a workshop — name it, pick the owner account, hit create workshop. This creates a private repo and clones it. Use this when the work has a real home: history, review, a shared repo.
- Create a mini-workshop (local) — name it, hit create mini-workshop. No repo, no account — just a local folder you point agents at. Perfect for scratch; graduate to full → later when it gets serious.
In step ②: pick the workshop (or type its folder), give the desk an optional name, choose the agent CLI, and hit open a desk. A terminal opens with the agent oriented and already working inside the workshop. Repeat to put more desks on the same problem.
A mini-workshop isn't listed in the "open a desk in" chips (it's not a full workshop yet) — open its first desk from its card's open a desk here.
With desks running, the board is your room at a glance:
- Each workshop card lists its desks: modality, model, output tokens, status, and when each was last seen. ↗ opens a desk's live console; × closes it.
- Signals are each desk's own self-report. A weak score or an ⚡ escalation rises into the TA (technical assistant) strip at the top of the card — the coordination layer flagging what needs a decision. Handled it? Dismiss it with its ×.
- Dormant desks (▶ to start) are desk folders with no live session yet.
- Flip simple → everything for the daily pulse, metrics, AIC, the shared bench, and the hands-up queue.
- Close a desk with its × — a live one winds down (it writes its journal first), then leaves the board. Bring it back later from closed desks (↻).
- Close a workshop with the × in its card header — hidden from the board but kept whole on disk. It tucks into closed workshops at the bottom; ↻ restores it. (Mini-workshops close the same way — or graduate them if they're worth keeping.)
Reach for it when:
- One artifact wants several perspectives (review + migration + tests, say).
- You're relaying between chat tabs and holding all the context in your head.
- The work spans sessions and you want the agents to remember.
Skip it when:
- It's a single quick question — the CLI alone is faster.
- There's genuinely one thing to do and one way to do it.
Screenshots use a throwaway demo dataset with fake names. Regenerate them with
tools/seed-demo.ps1 (see the script header), then point an instance at it via
the WORKSHOP_DIR / WORKSHOP_SESSION_ROOT environment variables.