Claude Code for New Zealand businesses.

What Claude Code is good at, and where it is not

Claude Code is an agent that does software work, not a chat window that suggests snippets. It searches the repository, plans, edits files, runs commands such as your test suite, and keeps going when something breaks, asking permission before changing files or running commands unless configured otherwise. As listed on 19 October 2026 it runs in the terminal, in VS Code and JetBrains extensions, in the Claude desktop app, on the web and mobile, and from Slack.

Where it earns its keep in New Zealand organisations:

• Explaining inherited code, such as the internal app a contractor built years ago that nobody has touched since.
• Tests, bug fixes and migrations where a test suite tells you whether the change worked.
• Small internal tools. Report generators, CSV clean-ups and glue code against systems with published APIs, such as Xero, ServiceM8 or HubSpot.

It is a poor fit for work nobody can check: plausible code is not correct code. It does not know your business rules (how you round GST, which customers are on special terms) unless you write them down, and it should not touch production systems or customer data without guard rails.

Levy Street builds with coding agents too. Levy Street’s open-source browser game, World of ClaudeCraft, had its first playable version built in about 48 hours in June 2026 with the Claude Fable 5 coding model. We also built SPAWN D, an open-source, invite-only, pre-alpha control plane for running CLI coding agents such as Claude Code and Codex on machines you own, from a browser or phone.

What Claude Code costs in New Zealand

Claude Code has no separate price. It comes with a paid Claude plan or bills per token through a Claude Console (API) account.

Claude plans as listed on 19 October 2026. From a New Zealand connection the plan prices use a bare “$” with the line “Prices include 15% GST”; Anthropic does not label the currency, and the currency of Max’s included API credits is also unlabelled. Enterprise is quoted in US$.

All plans share one usage pool across Claude chat and Claude Code, reset on a rolling five-hour window, and paid plans add weekly limits. On the API, prices per million tokens (input/output, USD) are Claude Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, Haiku 5.5 $0.10/$0.50 for prompts up to 100,000 tokens, and Fable 5.1 $10/$50. Fast mode for Opus 5.5 is $8/$40.

Illustrative arithmetic: a session that sends 1.5 million input tokens and receives 120,000 output tokens on Sonnet 5.5 costs about US$3.00 + US$1.20 = US$4.20 before caching discounts. Claude 4.7 and later models produce about 30% more tokens for the same text, so estimates built from older token counts come in low. Our guide to choosing between an API and a subscription shows when per-token billing beats a seat.

Where your code goes: data handling and the Privacy Act

Repositories hold more personal information than people expect: test fixtures, database dumps, logs and CSV exports. Treat what Claude Code reads as data sent to Anthropic.

• Training. Under commercial terms (Team, Enterprise, API and cloud platforms) Anthropic does not train on code or prompts unless the customer opts in. On Free, Pro and Max, data is used for training when the account’s model-improvement setting is on.
• Retention. Commercial use: 30 days, with zero data retention available to qualified Enterprise accounts. Consumer accounts: five years if training is allowed, 30 days if not. Transcripts sent with the /feedback command are kept for five years. Claude Fable requires 30-day retention for safety monitoring by default.
• On the laptop. Session transcripts sit in plaintext under ~/.claude/projects/ for 30 days by default, so device encryption matters.
• Where it is processed. Anthropic’s API offers only “us” or “global” inference. Through Amazon Bedrock, the “AU” cross-Region profile keeps Claude inference within Australia and New Zealand, and Claude Code supports Bedrock as a provider. The March 2026 launch list was older Claude models, so check what is offered now.

Under the Privacy Act 2020, a provider that processes information only on your behalf is treated as you (section 11), so sending it is not a disclosure and IPP 12’s overseas rules do not bite; you still answer for security under IPP 5. A provider that uses the information for its own purposes, such as training, also holds it, which makes it a disclosure under IPPs 11 and 12.

A starter project your team can run this week

Illustrative example: a 20-person Christchurch trades business exports invoices from its accounting system each month and spends half a day checking them in a spreadsheet. The goal is a small internal script, not a production system.

1. Pick a contained task. One script that reads a CSV export, normalises dates to DD/MM/YYYY, and flags lines where GST is not 15% of the GST-exclusive amount. Make a copy of one month’s export and replace customer names with synthetic values.
2. Install on a business account. Anthropic’s terminal install command is curl -fsSL https://claude.ai/install.sh | bash (its documentation covers Windows). Sign in with a Team seat or a Console account, not a personal plan.
3. Block secrets before the first prompt. Add deny rules to .claude/settings.json in the project so Claude Code cannot read .env files (see the configuration below). Deny rules take effect immediately for everyone who opens the project.
4. Write the house rules. Run /init to generate a starting CLAUDE.md, then add your own rules: NZ date format, cents rounding, “never overwrite input files”, “ask before installing packages”. Commit it so the whole team shares it.
5. Give it the task with tests first. Use the prompt below. Ask to see failing test cases before any fix.
6. Review the diff like a junior’s pull request. Run the tests yourself, read every changed line, and check two edge cases by hand. Our lesson on evaluating AI-generated code has a review checklist.
7. Measure it. Time next month’s check with the script. Keep it only if it saves time and the exceptions are right.

{
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(./.env.*)"
    ]
  }
}

Read CLAUDE.md first. Write clean_export.py that:
1. Reads every CSV in ./samples (synthetic data only).
2. Converts dates to DD/MM/YYYY and amounts to two decimal places.
3. Flags any line where GST is not 15% of the GST-exclusive amount,
   unless the tax column marks the line as zero-rated or exempt.
4. Writes cleaned.csv and exceptions.csv and never overwrites the input.
Write the tests first, run them, and show me the failing cases before
you change anything. Ask before installing any package.

The same pattern works with live data later through official connectors. Xero, for example, runs a remote MCP server, but its V1 tools write data without asking the user to confirm. Read our Xero integration guide before giving an agent write access.

Claude Code versus Cursor and the Claude app

• Choose Claude Code if your developers want an agent that takes a whole task and works beside the editor they already use (it has extensions for VS Code, JetBrains and Cursor).
• Choose Cursor if your team wants an AI-first editor and is comfortable with its 2026 ownership and pricing changes.
• Choose the Claude app for staff who write and analyse rather than code. The same Pro or Team seat covers both.
• Choose neither yet if nobody on the team can review code. Buying a finished product is often the better call; see custom AI versus off-the-shelf tools, or have us build it through custom AI development.

Where Claude Code goes wrong

• Tests that agree with the bug. If the agent writes both the code and the tests, both can share one misunderstanding. Write the acceptance cases yourself.
• More autonomy than you intended. Anthropic’s product page says Claude Code now runs in auto mode by default on Pro, Max and Team, so it can work longer while still catching risky commands. Keep deny rules, and run anything near production credentials in a separate environment.
• Limits mid-task. The five-hour window and weekly caps can stop a long job halfway. Ask for small commits so work is never stranded.
• Untrusted connectors. MCP servers extend what the agent can touch, and a malicious one can read whatever enters the session. See MCP for business systems.
• Unowned code. A script nobody understands becomes a liability when it breaks. Name an owner.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.