I got tired of watching 5 sub-agents work inside a single scrollback. So I gave each one a panel.

Every AI coding tool shows you the same thing: one conversation, one scrollback, one agent talking to you.

But that's not how the work actually happens.

It happens across multiple agents, multiple files, a browser, a terminal, and a dozen config files — and you're supposed to keep all of that in your head while one thread scrolls past.

I built DSI so you don't have to.

What it is

DSI is a Deepseek Harness web client built around a single idea: put everything on screen at once, in resizable, zoomable panels, so you can see the work instead of remembering it.

Multiple Agents and Sub-agents running side by side — not stacked in one thread. Alongside them: a workspace explorer, a real browser, a terminal, and the managers for prompts, skills, plugins and config.

You don't switch tabs. You resize panels.

Three things I'd point a skeptical developer at first

1. Sub-agents you can actually watch.

Sub-agents are usually invisible children of a main thread — you see their results, never their work. Here each one is a panel. You watch them run at the same time, and the sidebar keeps them findable: filter by name, jump to a day with the date picker, tuck sub-agents out of sight when the list gets busy. When you need full width, the whole rail folds away and comes back with one key.

2. The Glass browser — the agent shows its work instead of claiming it.

When an agent changes DSI itself — a plugin, a panel, a setting — it doesn't tell you "done." It runs the thing in the Glass panel and you watch it happen live. If a plugin breaks, you see exactly what broke instead of trusting a green test report. The browser is the receipt.

3. Prompts as muscle memory.

Type ?commit and your saved prompt appears, best match first — read it, edit a detail, send. Type ! and it runs immediately. A code review or a session summary becomes one short line instead of a paragraph you retype every day. The Prompt Manager (/dsi-prompts) is where they live, so the shortcuts always have something to find.

The honest trade-offs

(This section is why developers will trust the rest — keep it.)

• The Glass browser's brain runs as its own process by default. You can fold it into the page server with --playwright-mcp for a faster start — but then a brain failure can touch the page server itself. DSI prints a banner saying so. I'd rather warn you than surprise you.
• GLM (Z.AI) bootstrapping is optional. --zai is a convenience for people who have that account. No account? Leave the flag off — DSI uses whatever provider your harness already has.
• Running DSI in a VM with your screen on the host works, but you have to say so with GLASS_CHASSIS_URL. If the companion is down, the panel tells you honestly instead of silently opening a window you'll never see.

Try it

npx @deepseek-insight/dsi@latest dsh --sync --zai

Opens at http://127.0.0.1:5174. One Ctrl-C stops both DSI and DSH.

Repo → dsh-insight

Apache-2.0. Interface in English, Chinese, Indonesian, Spanish.

One question

What would you put in a panel that I haven't? I'm curious what the second screen looks like for other people — what's the tab you keep flipping to while an agent works?