Skip to content
Frank Huettner

Agentic AI in Business · Workshop II

Homework until 16 Nov, Part B: How Claude Code works here

On this page

7. What Claude Code can do, and what it cannot

Claude Code canClaude Code cannot
read and write in ~/kingo-pod/claudewrite anywhere else on your laptop
read kingo-pod: compose.yml, the guideschange kingo-pod, or run ./kingo
change your n8n workflows and data tables, your Langflow flows and your Postgres databasesee your keys: it gets stand-ins
look things up on the websee your own Claude Code settings, if you use Claude Code already

Sometimes, Claude Code wants to execute a command (which we don’t allow), and then it’ll tell you to run a command on its behalf in the terminal.

Quiz. Claude Code says that the sandbox blocks a command, and asks you to run it in your own terminal. Should you?

  • Yes: Claude Code needs it to go on
  • Only if I understand the command
  • No, never
Show the answer

Only run a command in your own terminal if you understand it. Claude Code cannot leave its sandbox, but a command you type does. When Claude Code says that the sandbox blocks it, that is the sandbox doing its job.

Quiz. In step 4 you gave kingo your n8n key. Does Claude Code know the key now?

  • Yes: it needs the key to reach n8n
  • No: it gets a stand-in
Show the answer

No. Claude Code gets only a stand-in. nono keeps the real key and puts it into each request that Claude Code sends to n8n. Hence, your key is not uploaded to the provider of the LLM, but Claude Code can still use it. The Langflow key and your OpenRouter key are kept the same way.

8. Claude Code always asks for permission: should we tell it to just do its job?

Claude Code always works in one of four modes. The mode decides who says yes before a command runs: you, the AI, or nobody. The line under your text shows the mode, for example ⏸ manual mode on.

The mode does not decide where Claude Code can write. The nono sandbox does that, as the table above shows: Claude Code cannot write into any other folder on your laptop, such as Documents or Desktop, with or without your permission. The mode only decides whether Claude Code asks you first.

To change the mode, press Shift+Tab. Each press takes you one mode further, in this order:

  1. ⏸︎ manual mode on: Claude Code asks you before it writes a file and before it runs a command. ./kingo claude always starts here.
  2. ⏵⏵ accept edits on: Claude Code writes files in its folder ~/kingo-pod/claude without asking. Before a command, it still asks you.
  3. ⏸︎ plan mode on: Claude Code changes nothing. It writes a plan and asks for your yes.
  4. ⏵⏵ auto mode on: the AI decides itself whether it needs to ask you for permission.

One more press of Shift+Tab, and you are back in manual mode.

Quiz. The line under your text says plan mode on. You ask Claude Code to rename a workflow. What happens?

  • Claude Code renames the workflow
  • Claude Code writes a plan and asks for your yes
Show the answer

Claude Code writes a plan and asks for your yes. In plan mode, it changes nothing before you agree to the plan.

⏵⏵ There is a fifth way, where nobody says yes or no: you start Claude Code with ./kingo yolo instead of ./kingo claude. Then Claude Code never asks for permission.

The difference, in five examples:

What Claude Code wants to do⏸︎ In manual mode⏵⏵ In accept edits mode⏵⏵ In auto mode⏵⏵ Started with ./kingo yolo
Write a file: save its findings in the file notes.md in its folder ~/kingo-pod/claudeClaude Code asks you for permission firstClaude Code writes the file without asking youClaude Code writes the file without asking youClaude Code writes the file without asking you
Run a command that only reads: list your n8n workflowsClaude Code asks you for permission firstClaude Code asks you for permission firstThe AI checks the command. If the AI judges it safe, Claude Code runs it without asking you. If not, the command does not runClaude Code runs the command without asking you, and the AI does not check it
Run a command that changes something: make a copy of a Langflow flowClaude Code asks you for permission firstClaude Code asks you for permission firstThe AI checks the command. If the AI judges it safe, Claude Code runs it without asking you. If not, the command does not runClaude Code runs the command without asking you, and the AI does not check it
Run a command that you said no to: you wrote “Do not change the original flow”, but Claude Code gets confused and wants to change the originalClaude Code asks you for permission first. If you say yes without reading, the command runsClaude Code asks you for permission first. If you say yes without reading, the command runsThe AI checks the command. It sees that you said no to this: the command does not run. But the AI can get it wrong too: then the command runsClaude Code runs the command without asking you, and the AI does not check it
Write a file outside its folder: save a file on your DesktopThe nono sandbox stops it, with or without permissionThe nono sandbox stops it, with or without permissionThe nono sandbox stops it, with or without permissionThe nono sandbox stops it, with or without permission

The fourth example shows the risk. An AI can get confused, and then it changes or deletes something that nobody asked for. This happened to a professor: he asked for new page numbers, and lost the lecture notes of a year.

Who stops such a command?

  • In manual mode, only you. You would have to read a long command and find the mistake in it.
  • In auto mode, the AI. It reads every command and compares it with what you wrote. That is better than a yes without reading, but the AI can get it wrong too.

So no mode makes you safe. Two other things do: the nono sandbox lets Claude Code write in its folder only, and every start of ./kingo claude saves a copy of your workflows and flows. There is no such copy of your Postgres database.

What does this mean for you? In this homework, Claude Code rarely writes files. It mostly runs commands to search, and to talk to n8n, Langflow and your Postgres database. In accept edits mode, Claude Code still asks you before each of these commands. The requests stop only in auto mode, or with ./kingo yolo.

Which mode should you use? Auto mode is okay. Claude Code runs inside the nono sandbox, so the sandbox protects your laptop in every mode. And every start of ./kingo claude saves your n8n workflows and your Langflow flows. In manual mode, you get many requests, and the commands are hard to read. So feel free to switch to auto mode: press Shift+Tab until the line under your text says auto mode on.

In manual mode, Claude Code asks before each command. Yes runs the command. No stops Claude Code: then tell it what to do instead, or type continue.

Quiz. You are in auto mode. Claude Code tells you that a request was blocked, and asks whether it may try again. It is the change you asked for. What do you do?

  • Type /exit and start again
  • Say yes, and name what it may do
  • Nothing: Claude Code finds another way
Show the answer

Say yes, and name what it may do, for example: Yes, retry saving the change to the copy. In auto mode, the AI may refuse a command that you asked for. Is it blocked again? Press Shift+Tab until the line under your text says manual mode on, and say it again.

9. You decide: ./kingo yolo starts Claude Code without any check

In each of the four modes, someone checks a command before it runs: you, or the AI. You can also start Claude Code with no check at all. To this end, you start it with this line:

./kingo yolo

The first time, Claude Code asks whether you accept this mode. Choose Yes, I accept.

./kingo claude ./kingo yolo claude (without kingo)
Your laptop isprotected by the nono sandbox and by Claude Code’s hesitationsprotected by the nono sandboxprotected only by Claude Code’s hesitations: no sandbox
Your workflows, flows and databaseyou see every change before it happensClaude Code changes them at once, mistakes tooout of reach at first, as Claude Code has no keys for them; but it can read every file you can, your keys too
Your user experiencemany requests for permissions, often the same onesClaude Code works on by itselfAnthropic’s standard Claude Code: your own Anthropic login and bill, and the usual requests for permissions

The requests for permission protect your project from Claude Code’s mistakes. If you let Claude Code work without asking, save your work first. This line saves your n8n workflows and your Langflow flows into ~/kingo-pod/backups/ (not your own tables in the database):

cd ~/kingo-pod && ./kingo backup

Quiz. Claude Code changed your n8n workflow WS2.3 - Triage New Mail, and now it does not work. Where is the version from before?

  • It is gone
  • In ~/kingo-pod/backups/flows/
  • In Claude Code's memory
Show the answer

In ~/kingo-pod/backups/flows/. Every start of ./kingo claude or ./kingo yolo saves each n8n workflow and each Langflow flow there as one JSON file, in a folder per start. The newest 10 starts stay. You can import such a file again in n8n or in Langflow.

10. What is different from a standard Claude Code, and where it is configured

The Claude Code you started with ./kingo claude is set up for the course.

Quiz. You are in the folder ~/kingo-pod/claude and type claude. Are you in the sandbox?

  • Yes: the folder belongs to kingo-pod
  • No
Show the answer

No. Then you get a standard Claude Code with Anthropic’s settings: no sandbox, your own ~/.claude, your own Anthropic login. Even in ~/kingo-pod/claude, a plain claude runs without the sandbox and without the keys. Here, you’re on your own, kid.

The course setup holds only when you start Claude Code with ./kingo claude or ./kingo yolo.

In detail: what ./kingo claude changes, and where

What is different from a Claude Code you install and start with claude, and the file where each part is configured:

A standard Claude CodeHere, started with ./kingo claudeConfigured for the course in
Startclaude, in any folder./kingo claude in the kingo-pod folder; ./kingo yolo is the same without requests for permission. Claude Code works in kingo-pod/claudekingo-pod/kingo, the script behind every ./kingo command. Before Claude Code starts, it checks that n8n and Langflow run, asks for your keys the first time, copies AGENTS.md, CLAUDE.md and .claude/settings.json into kingo-pod/claude/ the first time, and writes the nono profile. Then it starts Claude Code inside nono.
Model and billClaude models, with an Anthropic account or keyGLM 5.3 Flash through OpenRouter, paid with your OpenRouter keykingo-pod/claude/.claude/settings.json names the model, for every role (ANTHROPIC_MODEL and the lines ANTHROPIC_DEFAULT_…_MODEL). The nono profile sends Claude Code’s requests to OpenRouter instead of Anthropic, and puts your OpenRouter key into them. Claude Code does not know this model: kingo-pod/kingo tells it the context size to work with (CLAUDE_CODE_MAX_CONTEXT_TOKENS), 200,000 tokens.
Keyskept by Claude Code itselfkept in ~/.config/kingo/; Claude Code sees only stand-inskingo-pod/kingo asks for your OpenRouter and n8n keys, makes the Langflow key itself, and keeps them in ~/.config/kingo/, readable only by you. nono reads them from these files when it starts; they are never in an environment variable. The nono profile says, for OpenRouter, n8n and Langflow, which file holds the key and where nono’s proxy puts it into a request.
FilesClaude Code reads and writes wherever you may, after askingClaude Code writes only in kingo-pod/claude and ~/.kingo-claude, reads kingo-pod, and nothing else; the operating system enforces itThe nono profile ~/.config/nono/profiles/kingo-claude.json lists the folders Claude Code may read and write, and the few environment variables that reach Claude Code: other keys on your laptop do not. kingo writes it anew at every start from kingo-pod/setup/claude/nono-profile.json, so a change to it does not last. On top, kingo-pod/claude/.claude/settings.json forbids Claude Code to read .env files, ~/.ssh and ~/.aws.
SandboxClaude Code’s own sandbox, for commands only, if you turn it onnono, around all of Claude Code: its commands, and its own tools for files toothe nono profile. kingo-pod/claude/.claude/settings.json turns Claude Code’s own sandbox off (sandbox.enabled): nono covers more.
Settings and history~/.claude~/.kingo-claude; your own ~/.claude stays untouchedkingo-pod/kingo starts Claude Code with CLAUDE_CONFIG_DIR=~/.kingo-claude: Claude Code keeps its settings, sessions and history there. An Anthropic login stored there, kingo removes at every start: the course does not use it, and Claude Code could read it.
Instructionsa CLAUDE.md that you writeAGENTS.md: the part about kingo-pod and your team’s rules; CLAUDE.md only points to itkingo-pod/claude/AGENTS.md, yours to change, and kingo-pod/claude/CLAUDE.md, which holds only @AGENTS.md. settings.json makes Claude Code skip the CLAUDE.md of kingo-pod itself, which is for kingo-pod’s developers (claudeMdExcludes).
Memoryhidden in ~/.claude/projects/in kingo-pod/claude/memory/, where you can read itkingo-pod/claude/.claude/settings.json (autoMemoryDirectory)
What leaves the laptopthe model’s requests and the web pages Claude Code looks up, plus usage data, error reports and feedback to Anthropic; pages on claude.ai (artifacts) and Remote Control, if you use themthe model’s requests, to OpenRouter only, and the web pages Claude Code looks up; usage data, error reports, feedback, pages on claude.ai and Remote Control are offkingo-pod/kingo turns off pages on claude.ai, usage data, error reports and all traffic that is not needed, when it starts Claude Code. kingo-pod/claude/.claude/settings.json turns off the same, and also feedback (DISABLE_FEEDBACK_COMMAND), Remote Control (disableRemoteControl), pages on claude.ai (enableArtifact) and the check of each web address at Anthropic before Claude Code reads a page (skipWebFetchPreflight).
A way backnoneevery start saves each n8n workflow and Langflow flow as a JSON file in kingo-pod/backups/flows/, a folder per start; the newest 10 stay. Claude Code can read them but not change them. A file can be imported again in n8n or in Langflowkingo-pod/kingo, before it starts Claude Code. It needs psql.
Requests for permissionsince version 2.1.283, Claude Code starts in auto mode: a model checks each command, runs what it judges safe and blocks the rest, without askingClaude Code asks before a command (manual mode). Shift+Tab switches the mode, to auto mode too: there, GLM 5.3 Flash is the judge. ./kingo yolo turns the requests offkingo-pod/kingo starts Claude Code with --permission-mode default, and ./kingo yolo starts it with --dangerously-skip-permissions. The sandbox stays the same.
UpdatesClaude Code updates itselfno self-update: run the install line againkingo-pod/claude/.claude/settings.json: CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC also turns off the self-update.

Claude Code’s memory. Claude Code writes down for itself what it learns about your project, so that it still knows it in the next session: for example a name your team chose, or a mistake it should not make again. Usually this memory is hidden in Claude Code’s own folder. Here it lies in ~/kingo-pod/claude/memory/, so you can read what Claude Code remembers, and correct or delete it. The file MEMORY.md there lists everything it keeps.

Claude Code can explain all of this itself. Ask it:

How are you set up here, and what is different from a standard Claude Code?

Next: Recommended · Find out why your mail AI assistant missed the kimchi →