Keep what your coding agents make

The manual way

# copy it out before the worktree goes
mkdir -p ~/agent-output
cp /tmp/pricing-demo/index.html ~/agent-output/pricing-demo.html
cp .wisp/shots/phone.png ~/agent-output/

Then remember which folder holds what, and paste the path into the next prompt that needs it.

Save it

1. Start wisp.
   npxnpmcurlbrewDownload

   $npx wisp-agentsCopy

   macOS · Linux and Windows in beta · Node 20+ · MIT

2. From a reply. Pictures, videos, docs and code blocks have Save to My Stuff beside Copy and Download. It lights up once saved; click it again to take it out.
3. By asking. Save that to my stuff: the agent calls wisp_save, and its answer shows a card for what it saved.
4. By hand. Drop or paste files on the My Stuff page (the tray beside New task, or ⌘K → My Stuff).

Each thing is a copy in ~/.wisp/stuff, so it stays when the task, its worktree or /tmp goes. Cards preview the thing itself and say which task it came from. Search titles, notes, #tags and a doc’s text, filter by kind or star, and group by time or by task.

Use it in the next task

Type @ in any prompt or follow-up and pick a thing: the agent gets its file (or folder, or link), a doc’s text, and how to update it. Agents can also look for it themselves with wisp_stuff (use the logo from my stuff).

When an agent updates something you kept, it saves a new version in place: same card, title, star and tags, a v2 on it, and ⋯ → Restore brings back the one before.

Apps

For a demo, slides or a small tool, agents build a static web app and save its folder with a screenshot as its cover. Launch serves it on a port of its own and opens it inside wisp, with Reload, New tab, Stop and Close. It stops after 30 minutes idle. An app whose wisp-app.json says "db": true gets its own SQLite database that survives updates. To share one, see publish an app your agent built.

⋯ → Pin beside tasks makes an app a tab in the side panel, told which task is open; Pin on home puts it on the home page, for a board of your tasks or a PR queue. An app can read and start tasks when its wisp-app.json lists what it needs in "uses" (tasks:read, tasks:write, prs:read, crew:read, usage:read) and its page loads /_wisp/pane.js: you allow it once when you pin it and wisp checks every call. A task it starts or a follow-up it sends waits on wisp’s own card, and only your click on Send sends it; that access ends when the app’s code changes. wisp keeps its requests on its own address, but treat what it reads as visible to whoever wrote it. On home, panes sit side by side: drag between them or under them to resize, and double-click to reset. On a phone or in a narrow window each is a row that opens as a page, with Back. A spell can pin panes with its panes setting and lay them out with the split.panes part’s columns, widths ("2,1,1") and height.

The Agent board is one: every task across your harnesses, accounts and machines. Make a pane in My Stuff or ⌘K, or typing make me a pane that… in New task, has an agent build one: it writes the app into My Stuff, pins it, and the pane reloads in place each time the agent changes it. Pinning grants nothing, so press Allow… on its bar when it asks. The same dialog adds two starters: Agent board, every task in Working, Needs you, Review and Done columns across harnesses, accounts, machines and your crew; and Allowance board, each account’s usage windows, when they reset, and period by period how much went outside wisp and what 1% cost. A pane can load /_wisp/pane.css to look like wisp in light and dark.