ben merrill.
Sparky at a laptop: a small orange flame wearing goggles

Sparky Lab

A workshop for an AI that runs on a computer in my house. It plans, does the work in a real folder, and leaves a receipt for everything it touched.

A finished piece of work in Sparky Lab: the request, a four-step checklist all ticked, a short log of what was read, edited and run, and the answer

The receipt

Every run shows its work.

When a job finishes, it folds into a short log: the steps it set itself, the files it read, the commands it ran and whether they came back clean.

Every changed file is listed with its line count. One click opens the exact lines that were added or removed, so I never have to take its word for it.

The work log opened to a diff of index.html, with three added lines highlighted

Plan first

A plan I can read before anything changes.

In Plan mode it can look around, but it can't edit a file or run a command. What comes back is a card: the goal, the approach, the risks, the steps, and which files it expects to touch.

I can change the steps or just say what's wrong. Nothing starts until I press Start. Then the steps become a checklist that ticks as it goes.

A proposed plan card with a goal, a risk, five steps, four files and two checks, and a Start this plan button

Approvals

Risky commands wait for a yes.

Every command is sorted by what it could do. Reading and ordinary edits go ahead. Anything destructive stops and asks, and says why in plain words.

I can allow it once, allow that kind of thing for the rest of the conversation, or say no and tell it what to do instead. If I'm away, my phone rings.

A run paused on an approval card for the command rm -rf dist build, labelled Destructive, with Allow once and Deny buttons

The night shift

Stack the work, then walk away.

A conversation can be a queue. I add tasks in plain words, it takes them one at a time, and each one is its own run with its own receipt.

I can keep adding tasks while it works. The shift can run on a different model than the one I talk to during the day.

A shift queue with two tasks done, one being worked on and one waiting

Project notes

It learns the project, and shows me what it learned.

Each workspace has a short list of notes: how to run the checks, where things live, what bit us last time. They ride along with every new conversation in that folder.

When something it knows stops being true, it proposes the correction on a card. Replacing or retiring a note always waits for my tap, and the old wording is kept with the reason.

A Project notes card with two proposed updates: one new note to keep and one replacement for an older note

The library

Everything it made, in one place.

Every workspace is a shelf. Pages open as a live preview in a sandbox with no network, documents and code open as text, and anything can be edited right there.

Earlier versions of an edited file are kept, so a change I regret can be taken back.

The Library showing the files of a workspace on the left and a live preview of a small hiking club web page on the right

In your pocket

The same workshop on a small screen.

There's a native iPhone app too. These two pictures are the web app at phone width.

A new Work conversation on a phone-width screen, with Sparky waving above the message box
Start work from anywhere
The menu on a phone-width screen: Chat, Work and Library, with three workspaces and their conversations
Workspaces, one tap away

A workshop, not a chat box.

It lives at home

The models run on computers in my house. The conversations, the files and the notes stay on that box.

Three rooms

Chat is for talking. Work is for doing things in a folder. The Library is for reading what came out of it.

One folder at a time

Work happens inside the workspace I chose, and every command runs in a sandbox around it.

Pick the brain

Each conversation chooses its model. A bigger one for thinking, a second one on another machine for long shifts.

A door for Sparky

My personal assistant can hand a job to the Lab with a key that opens only that door. The result comes back as a report.

Long days stay quick

Long conversations are summarized before they get slow, and a line in the transcript says when that happened.