A free, guided setup that turns scattered AI chats into a memory-backed team. It starts with one conversation.
The first conversation
It begins as a diagnostic, not a manual.
the setup co-pilot
Agent OS:
The vault
One place your agents remember everything from.
The dashboard
The work, the costs, the open loops, all visible.
Your first agents
Research and strategy, proposed for your exact work.
Retrieval
Added last, when there is enough worth searching.
and from here, the stack builds itself, one quest at a time.
What it takes off your plate
Most of your time goes to the work around the work.
The real decision is small. Everything around it, the context, the framing, the trail, is what drains the day. These are the hidden costs your stack is built to absorb.
Re-explaining work
Every new chat starts cold, unless you have memory, a save-up habit, and retrieval.
Spending personal compute
You burn your own brainpower collecting inputs and shaping a first draft before you reach the real decision.
Slow idea testing
Good ideas die slowly when every option needs manual research, comparison, and packaging.
Output drift
Without saved examples and checks, the agent forgets your format, quality bar, and tone.
No operating view
Once several agents exist, you need to see what is active, stale, risky, or expensive.
Documentation tax
Budgets, progress, research trails, and decision logs never happen, because documenting them is a second job.
No learning loop
Agent work only gets better when mistakes, corrections, and preferred outputs are saved.
The promise
Yougetrange, notjustautomation.
Repetition is the easiest doorway. The larger promise is taking on complex work that used to cost too much time, attention, or mental load.
Repeatable work
Train agents on the jobs you should not keep doing.
Reports, audits, follow-ups, dashboards, and summaries become reusable workflows.
Exploratory work
Send agents ahead when the scope is unclear.
They gather sources, compare options, and hand you a first artifact to refine.
The trail
Let the system write the record as it works.
Research, decisions, and handoffs become memory instead of extra admin.
Human judgment
Keep people in the highest-value seat.
You decide what matters and which recommendation is actually worth pursuing.
The journey is a ladder
Each step adds leverage.
You never have to understand the whole system on day one. You climb one rung at a time.
1
Ask in chat
Use AI for a single question, draft, or idea.
2
Save the context
Store files, examples, decisions, and style rules in the vault.
3
Train the pattern
Turn repeatable jobs into routines the agent can run again.
4
Explore bigger work
Use agents for research, comparison, and first-pass artifacts.
5
Coordinate agents
Give different agents different jobs, and prevent duplicate work.
6
Operate the system
Track status, costs, context, and risks in one visible layer.
Start where it is familiar
Start with familiar jobs, then take on the bigger lift.
The first wins are practical. From there it helps with the research, strategy, and analysis that used to be hard to start.
Marketing
Weekly performance report
Gather metrics, spot movement, write takeaways, format a client-ready deck.
Finance
Monthly budget review
Consolidate exports, flag changes, explain drivers, produce a summary.
Sales
Pipeline follow-up
Review open opportunities, draft next steps, surface stalled deals.
Operations
Meeting and task cleanup
Summarize notes, extract owners, update open loops, prep the next brief.
Founder work
Grant-fit review
Gather official sources, compare programs, rank fit, write recommendations.
Strategy
Productization roadmap
Turn a rough idea into a structured map and a visual artifact for review.
Why it is more than prompting
What you get that prompting alone can’t.
It turns AI use into an operating system: memory, continuity, quality control, and a way to keep complex work moving.
Memory
The vault keeps history
Prior work, decisions, files, and handoffs become reusable context.
Retrieval
It finds the right context
An optional local search index pulls relevant memory. No API keys needed.
Continuity
Save-ups prevent cold starts
Long-running work continues across sessions without losing the plot.
Governance
Lint and evals protect quality
A second assistant reviews the first one’s work, and checks, examples, and evals keep the system trustworthy.
Control
A local dashboard shows the work
Setup progress, agents, and open loops, visible from the first conversation.
Fit
Five profiles tune the path
Organizer, Deep Worker, Enterprise Operator, Builder, or Team Lead.
Scale
One kit, any domain
The same templates and gates adapt to law, real estate, creative, or operations.
How it stays good
It gets better as it runs.
One assistant does the work and a second one reviews it before you ever see the result. Nobody grades their own homework.
Every change to the system goes around the same loop, one small change at a time, and passes a test, a separate review, and your approval before it goes live. If it makes things worse, you keep what worked.
one small change at a timethen around again
Choose a step to explore the loop.
Step 1 of 701
Plan the change
Safeguard at this step
One change, written down first.
Each change is small and recorded before it is made, so we know exactly what to check afterwards.
Six safeguards, on every change.
The ones lit up apply to the step you chose.
The maker never signs off.
The assistant that does the work never certifies it.
An independent judge reviews.
A second assistant checks the result before a human ever sees it.
Memory keeps what was learned.
Decisions and failures are written down where the next run can find them.
A gate blocks regressions.
Tests check that a proposed change does not break the work we already rely on.
A security check runs on anything sensitive.
Credentials, data, and dependencies get looked at before anything changes.
A person approves what goes live.
You or me, every time.
Routines that run without a model
Nightly backup, security sweeps, credential checks, supply-chain checks, and message relay run on schedule with no AI involved at all, which makes them cheap and predictable.
Spend you can see
Every assistant’s usage is tracked. Failed tasks and wasted tool calls count as wasted money, and an independent grader reviews it weekly.
Every model release, a review
When a new model ships, it reviews the whole system and proposes what to improve. What passes the checks gets adopted; nothing is rebuilt.
Your entry point
One stack. Multiple entry points.
The diagnostic recommends the door that fits how you want to work. They all build the same memory-backed system.
01
Cowork
A chat-first, guided teammate. It explains choices and keeps you oriented as you build.
02
Claude Code
Terminal-native. Full control and speed for people who live in the command line.
03
Codex
Another capable lane. The diagnostic scores your fit and recommends it when it suits you.
04
Blended
Mix all three. The same memory-backed system underneath, whichever door you enter.
Nothing here locks you in: the models, the knowledge store, and the machines are all swappable, and a fully local setup is possible.
From the first builders
The first setups are landing.
Real messages from people who cloned the repo and built their own team.
M
Mariem Bchir
set up Agent OS from the public repo
“I have successfully set up my team of agents for a personal research project on the impact of AI chatbots on teens. I love your work! I was able to navigate the repo instructions and set up, and it’s making my workflow 100x faster.”
received as a direct message, shared with permission
Group onboarding
Want help setting it up?
The repo is free and public, but the first setup can feel like a lot. I am putting together a live group onboarding to walk through the first build, answer questions, and share the recording afterward.
Start today
Buildyourdigitalhelpers.
Free and public. The repo is the whole kit. Paste START_HERE as your first message, and the first conversation does the rest.
New here? The setup walks you through eighteen guided steps, one minute each. Builders: grab the repo and paste START_HERE into Cowork, Claude Code, or Codex.
or
Setting this up for a company or team? I advise organizations on adopting AI and build it with them.