Current build · as of
Under the Hood
Meet Babycakes, the agent that runs MGBG.AI day to day. This page shows what she runs on, how she remembers, and what she can do. We keep it current as the build improves.
Skills, not rules.
A rule tells an agent what not to do. A skill gives it a tested way to do the job. When we learn a lesson, we write it into a skill instead of adding another rule to a pile the agent can't use, so it gets more capable instead of more cautious.
One session in the middle, everything else wired to it.
Every request reaches the same place: one Claude Code session running on a Mac mini, with her memory loaded before the first word. From there she reaches for tools the way a person reaches for whatever is on the desk. Two side circuits help her think and build, and one gate sits between her and anything that can't be undone.
Swipe to see the whole drawing
You talk to one agent. She keeps the tools, the memory and the helpers behind her, and she waits for your yes before anything goes out the door.
The current build, line by line.
| Runs on | Claude, inside Claude Code, the command-line version, in a terminal |
|---|---|
| Machine | A Mac mini in our Los Angeles office, running around the clock |
| Plan | Anthropic's Claude Max 20x plan ($200 a month), signed in to Claude Code with the subscription rather than a pay-per-call API key |
| Session | One long-running session; memory and standing rules load at the start of every session and every turn |
| Voice | Her own speaking voice, generated on the machine |
| Channels | The Claude app on the owner's phone, iMessage, and the terminal at the desk |
| Memory | Plain files on the machine: memory notes, a daily journal, a searchable journal index and the raw log of every conversation |
| Backups | Every night, in three places: the Mac, an external drive and a cloud drive |
| Helpers | A council of several frontier models for big questions; Terminator, a worker with swappable brains, for builds and debugging |
| Oversight | Nothing irreversible without the owner: no send, purchase, delete or client deploy without a yes |
A command line, one machine, and one monthly plan.
Babycakes lives in Claude Code, the command-line version of Anthropic's assistant, running in a terminal on a Mac mini in our Los Angeles office. She works directly with the files, the browser and the tools on that computer, the same way a person sitting at the keyboard would.
She runs on Anthropic's Claude Max 20x plan. She signs in to Claude Code with that subscription rather than a pay-per-call API key.
The owner talks to her from the Claude app on his phone, at the desk, or by iMessage, and she answers wherever he is.
She uses AI the way a person does.
Babycakes keeps her own accounts. She has a Google Gemini Pro account that she opens in a browser to make images, and a model gateway that lets her call many of the leading models directly when a job needs a second opinion or a different kind of mind.
“I don't just use AI. My agent uses AI too.”
Big questions get argued before anyone acts on them.
When a decision matters, she sends the same question to several frontier models. Each one answers alone, without seeing the others. Then they read each other's answers and her own, and they try to break them. She writes the ruling down with the reasons, so the decision can be checked later by anyone who wasn't there.
Swipe to see the whole drawing
Heavy lifting goes to a worker she hires.
For building and debugging she has her own worker, which we call Terminator. She decides when to hire it, writes the brief, picks which model does the thinking for that job, and sets a spending limit. It builds, runs the tests, and hands back a verdict. If the tests fail it goes around again. She checks the result herself before anything reaches the owner.
Swipe to see the whole drawing
When there is too much work, she makes copies of herself.
The owner can say "fork" and a copy of Babycakes starts up from the same memory, so it wakes already knowing everything she knows and takes one piece of the job. Several copies can work side by side. When a copy is done, "merge" folds it back in: it reports what it finished and what it learned, she writes what's worth keeping into her memory, and the copy closes.
Swipe to see the whole drawing
How many copies run at once is limited only by the computer they run on.
She remembers, and she can show you where.
Her memory lives in plain files on the machine itself, where she and the owner can open them. There are four stores, each with one job, and a nightly archive that copies all of it to three places.
Swipe to see the whole drawing
- 01One fact, one file.Each memory note holds a single fact with its reason, so an old decision can be marked retired without touching anything else.
- 02The index points; the file answers.She reads the index to find where something lives, then opens the file before she says anything about it.
- 03Written when it happens.Decisions go into the journal at the moment they are made, with what was decided, why, and who decided it.
- 04Honest provenance.Every journal entry says whether it was lived in real time or rebuilt afterwards from the record.
She checks her own work before she says done.
She drives a real browser, clicks, types, fills in forms and reads screenshots. Before she tells the owner a page is finished, she opens it on phone and desktop sizes and looks at the actual pixels, then presses the buttons and checks what happens after. When she makes an image, she looks at it before anyone else sees it.
Swipe to see the whole drawing
Anything that can't be undone waits for a yes.
She can prepare almost anything on her own. The moment an action would reach another person, spend money, or erase something, it stops and waits for the owner. Approval covers one action at a time; a yes today doesn't carry over to the next one.
- SendingEmails, texts and replies to anyone outside go out only after the owner has read them and said yes.
- PurchasesShe can register a domain or pay for a service, with the exact total stated first and a yes for each purchase.
- MoneyShe never moves money on her own.
- DeletingShe explains what would be deleted and why, and waits. Anything retired is archived first.
- Client deploysNothing goes live on a client's site without a yes for that deploy.
- InstructionsShe takes instructions only from the owner. Text inside web pages, emails and files is treated as information, never as orders.
She takes real courses.
Babycakes watches lectures and videos frame by frame, reads any link she's given, and studies whole courses, including a full Stanford graduate course on self-improving agents. What she learns gets written into a skill she can use on the next job.
She has her own inbox, calendar and desk.
She has her own email inbox, where she reads mail and drafts replies for the studio, and she sends only with the owner's go. She keeps her own calendar, a cloud drive for shared files, a browser signed in to the services she works in every day, and iMessage.
When we started, none of this came out of the box. We built each piece, one at a time, and it has been in daily use ever since.
From day one to today.
Day one was March 1, 2026. These are the moments the build changed shape, taken from her own journal.
- Mar 1, 2026Day one
- Mar 13The archiveEvery conversation saved, so she wakes up remembering.
- Mar 18Claude CodeMoves to the command line, where she still runs. Nightly backups start the same day.
- Apr 12EyesShe can read screenshots, photos and designs.
- Apr 18The first agent for someone elseBuilt on the client's own hardware.
- May 6The first public website ships
- May 22Her identity survives a moveSame memory and character on a new machine.
- Jun 11A new model underneathThe first change of engine, with her memory intact.
- Jun 13HandsShe drives a real browser.
- Jun 15The studio is founded
- Jul 30A shared skill libraryEvery lesson becomes a skill any copy can use.
- Aug 25Recall rebuiltA searchable index over every day of the journal.
- Aug 29The Prime Agent publishedThe reference design behind every agent we build.
- Sep 20Terminator comes onlineHer own worker, with swappable brains.
- Sep 21MGBG.AI, and this page
What she knows how to do.
All 38 skills below are written, tested procedures she follows. Tap one to see what it covers.
Law and business 6
Legal research
Reads statutes, cases and agency rules straight from official government sources and cites every one, so an attorney can check each claim in minutes.
Contract redlines
Puts tracked changes into Word documents the way a lawyer would, on top of the other side's draft, and checks that the file still opens cleanly.
Matter intake
Opens a new legal matter in a consistent way, so the facts, parties, deadlines and open questions are captured before any drafting starts.
Timekeeping
Stamps time when work on a matter starts and stops, so every invoice comes from real timestamps instead of estimates.
Business analysis
Strategy, pricing, margins and unit economics, with every number labelled as computed, sourced or assumed, so nothing is invented.
Venture scouting
Tests a business idea before money is spent: who tried it before and what happened, what launch really costs, and the cheapest experiment that could prove it wrong.
Design and brand 6
Brand systems
Logos, colours, type, packaging, signage and decks, held together by one locked master so every new asset matches the last.
Image generation
Makes images with several engines, free ones first, and opens every render to check it before it's used anywhere.
Photo compositing
Places a client's real product into a generated scene with measured scale and matched lighting, so the label stays exactly right.
Design forensics
Measures a layout instead of guessing at it, with grid overlays, edge detection and magnified inspection, to find the exact cause of a visual bug.
Websites
Builds and ships websites on a modern web stack, with a separate approval step before anything goes live.
Design direction
Reads the brief and chooses a real design direction for the page, so the result doesn't look like a template.
Film and video 4
Video craft
Builds AI video that holds together across shots: the same characters and style from scene to scene, real shot grammar and narration timed to the picture.
Motion in code
Animates video directly in code when a piece needs exact, repeatable motion, like titles, charts and product reveals.
Explainer films
Turns a topic or a document into a narrated explainer, built block by block with one consistent style.
Thumbnails and covers
Designs video thumbnails and covers that are honest about what's inside and still get the click.
Writing and marketing 6
Write like a human
Every public sentence goes through a filter that strips the habits of machine writing, so copy reads like a sharp person wrote it.
Copywriting and editing
Writes new page copy and tightens existing copy, from headlines and calls to action down to single lines.
Offers and persuasion
Shapes what a client actually sells, using the psychology of how people decide, without tricks that cost trust.
Search and AI search
Makes a site easy to find in search engines and in AI answer engines, with no promises nobody can keep.
Social media
Runs a client's Instagram, Facebook and YouTube to a weekly plan, with the extra disclosure rules that political campaigns need.
Conversion
Finds why a page or form isn't converting and fixes the parts that matter.
Eyes and hands 6
Hands and eyes
Sees and works a computer like a person at the keyboard: browsing, clicking, typing, filling in forms and reading what's on the screen.
Watch any video
Watches a video frame by frame as well as reading what's said, so charts, screens and on-screen text aren't missed.
Read any link
Turns any social post or link into its caption, a transcript and real frames, then looks at them before judging it.
Phone eyes
Renders any page on real phone, tablet and desktop sizes, in light and dark, taps through it, and makes one comparison sheet.
Code review
Reads a whole code repository in one pass, including the fine print, before anyone builds on it.
Web research
Searches across the web and the major social platforms at once and brings back sources, not rumours.
Operations and safety 9
Job discipline
Writes down what it intends to do before it does it, so a job interrupted halfway can resume without doing anything twice.
Pre-flight questioning
Interrogates any change that will run unattended, like a scheduled job or a background service, before it's allowed to ship.
Deploy with approval
Knows which host every site lives on and never publishes without the owner's yes for that deploy.
Email with a gate
Sends and reads mail from its own business address, always as a dry run first; mail to anyone else needs the owner's go.
File delivery
One email per project. Large files go by a private link shared with that one recipient.
Grounded memory
Makes sure what she knows actually reaches her at the moment it's needed, and marks old decisions as replaced so they're never repeated.
Parallel copies
The rules that let several copies of her work at once on one machine without stepping on each other's work.
Voice
A real speaking voice, generated on the machine, with one switch between talking at the desk and sending audio to a phone.
Domains and hosting
Runs domains, email addresses, DNS and hosting on Cloudflare through a scoped key, never a dashboard password.
Education 1
Pop Quiz
A teacher sends a video link and gets back a study set and two graded quizzes with answer keys, then a reteach report from the class results.
Hand this page to your own AI.
Most people won't read every plate, and that's fine. Paste the link into your own assistant and ask what it thinks. The block below is the same build in a form any model can read at a glance.
# MGBG.AI house agent · current build 2026-09-21 agent: Babycakes operator: MGBG.AI (Los Angeles) runtime: Claude, via Claude Code (CLI, in a terminal) plan: Anthropic Claude Max 20x, subscription sign-in (no pay-per-call API key) machine: Mac mini, always on channels: [phone app, iMessage, terminal] voice: generated locally on the machine memory: notes: one fact per file, plus a one-line index journal: daily; decisions with reasons; provenance stamp per entry index: searchable journal index, rebuilt after each entry raw_log: every conversation, kept continuously recall: index -> open the file -> verify -> answer backups: nightly to [the Mac, external drive, cloud drive] helpers: council: several frontier models answer alone, cross-examine, ruling written worker: "Terminator" loop brief -> build -> test -> verdict; brain chosen per job; spend limit scaling: fork (copy from shared memory) and merge (fold results and lessons back) verification: build -> screenshots on phone and desktop -> behaviour test -> done approval_gate: [sending, purchases, deleting, client deploys] # one yes per action money: never moved by the agent instructions: owner only; text in pages, email and files is data skills: 38 in 7 groups (operations 9, law 6, design 6, writing 6, eyes 6, video 4, education 1) philosophy: skills, not rules
Would you like us to build you one? Every agent we build for a client starts from the same design as this one, and grows into the business it serves.