How to use HyperTeams Connect and the HyperTeams app on your machine.
Do not read it end to end — pick the path that matches your role.
The right order differs by role. The management track is written so you can skip the technical chapters without getting stuck.
Executives and programme leads
Deciding budget and priorities · 112 min
Then read “The adoption process” in order. You can skip the technical parts without getting stuck.
Practitioners
Using it every day · 121 min
Then go to “Getting started” for whichever product you use — HyperTeams or Connect.
No technical background
Getting stuck on the technical words · 58 min
Then go to “AI foundations”. The install and API chapters in the product parts will now read straight through.
Engineers and technical leads
Building, integrating, securing · 93 min
Then continue to “Hyper-agile” and HyperTeams “Extending it”.
From AlphaGo to now
Why nothing changed for six years after 2016, and what changed in 2022.
What AI is
This is not about robots. Start with how it differs from search.
What gets better
Five things that actually change in a working day — with real inputs and outputs.
Your first tryPractice
Three first conversations you can copy as-is. Five minutes.
Getting the answer you wanted
Five techniques that change the result for the same request — bad example, good example.
Ask for the same thing three timesPractice
Not a chapter to read — a chapter to do. Add one technique at a time and watch the answer change.
Things to watch out for
Made-up facts · stale information · company data. When to trust and when to check.
Adoption is not transformation
Why the same tool moves the numbers at one company and not the other.
Why now
Why things that did not work a few years ago work now — three changes at once.
How transformations fail
Three typical failures, and how many months in you can spot each one.
Score your own organisation's readinessPractice
Not a chapter to read — a chapter to do. Put your own organisation against the three failure modes.
Three questions for choosing work
There are always more candidates than you can do. Criteria for what to drop.
Quality
Does the customer-facing product get better — effects arrive late and last long.
Cost
The fastest area to prove. Which makes it the best first project.
Revenue and margin
The largest impact. And the most prerequisites.
Sequencing
Impact order and starting order are opposites. Why, and how to decide.
Actually score your candidatesPractice
Not a chapter to read — a chapter to do. Write six candidates and throw half away.
The roadmap
Four phases, and the gate that decides whether you may move on.
Phase 1 — define the problem
Define a specific business problem AI can solve.
Phase 2 — data and system assessment
Confirm "can it be built" before building it. Skip this and you stall mid-POC.
Regulation and compliance
Korea's AI Framework Act has been in force since 22 January 2026. Start by checking whether it applies to you.
Phase 3 — design and run a POC
The most important phase. Its purpose is not to succeed but to decide.
Design a POC on one pagePractice
Not a chapter to read — a chapter to do. It is not finished until the stop conditions are written.
Phase 4 — production rollout
Scaling is not widening access. It is embedding.
Measuring and reporting results
"It feels better" does not get funded. Miss the pre-deployment baseline and you can never prove it.
Design a baseline yourselfPractice
Not a chapter to read — a chapter to do. Measure today or never.
Audit trail and observability
You have to be able to answer "why did the agent do that?" What to record, and what never to.
Change management
The top cause of failure is people, not technology. The half the roadmap does not write down.
What your computer is doing right now
Storage, desk, worker — three things to tell apart. It is also why "restart it" works.
Files, folders, paths
Read one path and neither working folders nor Drive will trip you up.
The terminal and commands
Why the black window is still here, and what "command not found" really means.
Programs and languages — Python, JavaScript, Node
Why you are told to install Node, and why the examples are always Python.
Networks and communication
IP, ports, domains. This is where localhost:27777 finally makes sense.
Servers, clients, the cloud
A server is not a special machine. And the cloud is someone else's computer.
Requests and responses — HTTP
One round trip is the whole thing. Including why 401 and 404 are different problems.
What an API is
A counter opened for programs. And what changes the moment you use one.
Take one request apartPractice
A browser is all you need. See a request, a response and a status code for real.
Shapes of data — JSON, CSV, documents
Why some files an AI just reads, and others need converting first.
What a database is
Tables, rows, columns, keys. And why a spreadsheet stops working.
Accounts, passwords, API keys, permissions
Who are you and what may you do are two separate questions.
What a language model actually does
How "guess the next word" turns into summarising and translating.
Why it is wrong so plausibly
Hallucination is not a bug; it comes from the structure. So you manage it rather than remove it.
Tool use
Giving the model hands — it stops only saying what it knows and starts looking things up.
Finding and attaching sources
The most reliable way to stop it pretending to know. Search first, then answer with citations.
Try attaching sourcesPractice
Not a chapter to read — a chapter to do. Thirty minutes.
Which kind of learning?
The answer to "can't we just train it on our data?" — there are four kinds, and one of them is the expensive one.
Agents — the loop that runs itself
The difference between using a tool once and taking a goal and repeating until done.
The context window
Everything the model is currently looking at. What "memory" really is.
When context gets pushed out
Why it acts as though it forgot your instructions, and the habits that prevent it.
Push context out on purposePractice
Not a chapter to read — a chapter to do. Manufacture the moment the beginning falls off.
Memory that outlives the session
Close the window and it is all gone. Three layers, and what belongs in each.
The four parts of a good instruction
Goal · scope · done condition · context. What goes wrong when each is missing.
Rewriting instructionsPractice
Five bad instructions to fix yourself. Write your version before looking at the answer.
Evaluation (evals)
Knowing whether you are doing well from numbers rather than instinct. Mandatory at organisational scale.
Choosing a model, and cost
Run everything on the best one and you pay 3-5x. What goes where, and the three ways to save.
MCP — why a standard was needed
When attaching 3 tools to 3 agents meant building 9 integrations.
What an MCP server provides
Tools · resources · prompts. In practice it is mostly tools.
When you have too many tools
Why attaching 200 hurts accuracy too, and the mechanisms for narrowing them down.
When what it reads becomes a command
Prompt injection. Instructions planted in a web page, document, or email becoming the agent's orders.
Tools are permissions
Attaching an MCP server is granting the agent a permission.
Audit your tool permissionsPractice
Not a chapter to read — a chapter to do. Walk the tools you have enabled right now.
What agile solved and what still blocks
What the two-week sprint was actually sized for. And what is inside it.
What hyper-agile changes
Not shorter sprints. A different thing doing the coordinating.
Micro-sprints
The point is independence, not size. Why splitting alone is not enough.
Try splitting your own backlogPractice
Not a chapter to read — a chapter to do. Split one live work item in forty minutes.
IntentOps
The developer's output changes from code to intent. So what do they own?
Fluid software
The definition of legacy changes. And what you must protect moves.
Map of the four processes
Where analysis, design, build and test went. A map before the next four chapters.
Process 1 — capturing intent
A PM agent asks back. Ambiguity left here does not resolve itself.
Process 2 — agent orchestration
Not a human team's hierarchy transplanted, but roles running concurrently.
Process 3 — automatic validation
Where the self-healing loop runs. And where this methodology most often collapses.
Design your validation gatesPractice
Not a chapter to read — a chapter to do. Write the conditions under which it can ship unwatched.
Process 4 — immediate delivery
Deployment is not the end but the start of the next intent.
Guardrails and a pre-adoption checklist
Not devices that slow you down — conditions that let you survive the speed.
What HyperTeams is
A control room for directing Claude Code on your own machine from a web screen.
Prerequisites
Claude Code and Node.js. Why each is needed and how to check.
Installing
One line. And what it asks you during the install.
First run and connecting
One word to start, a browser to get in. And how to stop it.
Check your first connectionPractice
Not a chapter to read — a chapter to do. Walk through each place people get stuck, before you are stuck.
Registering a working directory
One folder, one card. How to choose them and how to split them.
Your first taskPractice
A first task you can follow along with. Start in plan-only mode.
Deciding how much to hand over
The four autonomy levels and when to use each.
The terminal
A command window that stays open in a working directory. It survives closing the screen.
Scheduled runs
Handing repeating work to the clock. Designing for when nobody is watching.
Access from anywhere
A tunnel gives you a fixed address. Lock it before you open it.
Command reference
Start · stop · restart · setup · upgrade · uninstall.
Troubleshooting
A chapter to look up by symptom. Starting with how to read the log.
Recover a stalled jobPractice
Not a chapter to read — a chapter to do. Break it on purpose and walk the recovery in order.
The two API surfaces
Both default to off. Start with what opens and what does not.
Directing work over REST
Hand work to your machine from another system. Starting with the missing idempotency key.
Drive a job over RESTPractice
Not a chapter to read — a chapter to do. Create the duplicate run yourself.
Getting models — downloading and running them
The next chapter says "your machine's model answers". This is where that model comes from.
Your machine as a model server
It is OpenAI compatible, so the SDK stays. Only the base address changes.
Working with audio and images
Text to speech, speech to text, image generation. And cloned voices, which OpenAI has no spec for.
Building pipelines with schedules
There are two kinds of schedule — hand it to an agent, or run a command.
Build one pipeline end to endPractice
Not a chapter to read — a chapter to do. Chain three stages and break the middle one.
Capturing procedures as skills
A unit that outlives an instruction. The folder is the state.
Turning documents into material
PDF, Word, Excel and PowerPoint into searchable copies. And how git fits in.
Case study — a content factory
The previous six chapters joined into one. A production pipeline that runs daily.
Collecting with a browser — Crawlee + stealth
How to pull data from pages that only open in a real browser. The wiring is three lines; the hard parts are on either side of it.
Getting material out of video — yt-dlp + speech-to-text
Turning YouTube, Instagram and TikTok videos into material you can read. Most of the time you never need the video file at all.
Korean official documents — hwp and hwpx
Reading, generating and filling Hangul documents. Without Hancom Office, and without Windows.
What Connect is
An AI workspace a team shares. How it differs from HyperTeams.
Workspaces
Everything lives under one. How to split them is your first decision.
Threads
One job, one conversation. It keeps running after you close the window.
The work board — conversations as work
Open a group and you get cards. What needs you, what is running, what nobody has touched.
Run the work board for a dayPractice
Not a chapter to read — a chapter to do. Treat three conversations as work items.
Creating an account and workspace
From sign-up to a working workspace. Follow along with the screens.
Your first conversationPractice
Three you can copy as-is. Including checking a tool call.
Slash commands — /every
Turn what you are typing into a repeat. And retrying is no longer a command — it is the default.
What agents are
Five kinds and when to use each. How to read the list screen.
Creating a sub-agentPractice
How to write the instructions. One real example from start to finish.
Connecting an MCP server
There are two directions. Outside-in, and inside-out.
Connecting a database
PostgreSQL, MySQL, SQLite, MongoDB as tools. Start with a read-only account.
The playground
Call tools directly. The place to check before you rely on them.
Turning tools on and off
Turning off tools you do not use raises accuracy.
Drive
The workspace's file store. What you upload becomes material for answers.
Connecting a system
Registering your machine to a workspace. The token is a password.
Managing systems — slots and addresses
The first machine is free. What slots, addresses and tags each are.
Scheduled — stack it up, hand it to the clock
The queue and schedules are one screen now. And one word — scheduled.
Set one schedule runningPractice
Not a chapter to read — a chapter to do. Design for the times nobody is watching.
Channels — the ways in
Four ways to reach an agent without using the screen.
Email agents
One address becomes a contact. Send mail, get a reply.
Chat widget
A support window embedded in your site. Always restrict the domain.
Members and roles
Three roles and what each can do. How many admins to have.
Groups and departments
Groups group conversations, departments group people. Grant permissions to departments.
Pending approval and operator rights
Requests from people without run rights stop before they execute. Who releases them.
Actually divide up the permissionsPractice
Not a chapter to read — a chapter to do. Grant to groups, not to names.
Usage and credits
Look at the distribution, not the total. That is where the leak is.
Plans and tiers
Credits are not bought and topped up — a monthly plan issues them. What each tier unlocks.
Calling it via API key
Talking to a workspace from outside. Keys are harder to retire than to issue.
Getting notifications
Your phone can tell you while the app is closed. The switches live inside the header bell.
Glossary
Register the abbreviations only your company uses. Searches start finding what they used to miss.
One page for security review
The questions an internal review asks. Where the boundaries are and what locks each one.
Data masking
Personal data is swapped for placeholders before it reaches the model. Replaced, not deleted.
Audit log
Who asked what, and which document the answer came from.
Reconstruct a week from the audit logPractice
Not a chapter to read — a chapter to do. Answer four questions from the log alone.
SSO and department sync
Sign in with company accounts, and map their groups to departments.
Troubleshooting
What to check first when it doesn't work, symptom by symptom.
HyperTeams