For engineering and business teams
Claude Code in your company,
for the whole team
Your team has installed Claude Code, and everyone experiments on their own. We set up shared rules, skills and connectors to your systems for everyone, and your IT sets permissions centrally. We bring what has worked in our own operation.
- ›Take the next item from TODO.md.
- Reads CLAUDE.md and the rules
- Two subagents review copy and code
- Type check passed
- Build passed
- SEO check passed
- Pull request opened, waiting for approval
What is Claude Code, and is it worth it for your company?
Claude Code is an agent harness from Anthropic: the software around the Claude models that turns them into an agent for your code. The agent reads the codebase, edits files and runs commands. Claude Code pays off as soon as someone in your company builds software or tools. The skills and connectors you create along the way also load in chat and in Cowork, the Claude app for work that involves no code. That is why business teams work with it too.
Your developers are not the only ones who use it
- Leadership
- Leadership sets one goal per quarter and one workflow to measure it by. Leadership also makes two decisions: which plan to choose and which data Claude may see.
- Engineering
- The engineering team works with Claude Code in the terminal or in their code editor, with the CLAUDE.md in the repository, with hooks and a review step before every merge. Claude helps review pull requests.
- Business teams
- Sales, purchasing and accounting use the same skills and connectors in chat and in Cowork, with your rules and without a terminal. Anyone who wants to build small tools themselves uses Claude Code in the desktop app.
- IT and data protection
- IT sets managed settings, approves connectors and plugins, sets up single sign-on and sees usage in a dashboard. These settings then apply on every machine.
You have the tool. Your day-to-day hasn't changed.
You've heard about Claude Code: from your own team, from a competitor or in the press. Someone installed it. That is where the real work starts.
- Everyone configures it on their own. What works for one person, the next one never hears about.
- Business teams watch from the sidelines, because nobody set up skills and connectors for them.
- The repository has no rules, so every session starts from zero.
- Nobody has decided what the tool may access and what it may not.
- The results look plausible. Nobody has checked them.
- Nobody knows exactly what it costs to run per month.
What Claude Code needs in a company: nine building blocks.
| Building block | What it is | What it means for you |
|---|---|---|
| CLAUDE.mdper project, per person, company-wide | A Markdown file with instructions that Claude Code reads at the start of every session. It exists per project, per user and as a centrally managed file for the whole organization. Anthropic recommends keeping each file under 200 lines. | Every session knows your rules instead of starting from zero. The file lives in the Git repository, and the team maintains it together. |
| SkillsClaude Code, Claude, Cowork | A SKILL.md with instructions for one specific task. Claude loads the full content only when that task comes up. Skills follow the open Agent Skills standard and load in Claude Code, in chat and in Cowork. | Checking a quote, preparing a release, writing in your tone of voice: a workflow runs the same way every time, no matter who starts it. |
| Connectors and MCP serversaccess to your systems | MCP is an open standard that lets Claude access tools, databases and APIs. A .mcp.json file in the repository gives the whole team the same connections. On Team and Enterprise, only admins add connectors. | Claude reads the data where it lives, so nobody copies it back and forth anymore. Your IT decides which connections are allowed. |
| Plugins and marketplacedistributed from a company repository | A plugin bundles skills, connectors, subagents and hooks. A marketplace is a Git repository that lists plugins. On Team and Enterprise, Owners decide whether a plugin is available, installed by default or required. | One repository reaches both groups: developers in Claude Code, everyone else in chat and in Cowork. Nobody has to build their setup alone. |
| Subagentssubtasks with their own context | Specialized helpers with their own context window, their own task and their own permissions. They return only a summary, and routine work can go to a cheaper model such as Haiku. | Research, review and implementation run separately. The main chat stays focused, and a reviewer looks at the work with fresh eyes. |
| Hooksrules that always apply | Commands that Claude Code runs automatically at fixed points, for example before or after every tool call. Anthropic calls this deterministic control: the action always happens instead of the model deciding. | What has to happen doesn't depend on the model remembering it: formatting after every change, a stop for dangerous commands. |
| Permissions and managed settingsset centrally by IT | Rules that say which commands and which kinds of access are blocked, need confirmation first or are allowed, plus a sandbox for shell commands. The organization pushes managed settings to every machine, on Team and Enterprise straight from the admin console. No user and no project can override them. | Your IT decides once for every machine, instead of everyone configuring their own. |
| Surfacesterminal, IDE, desktop, web, phone | Claude Code runs in the terminal, in VS Code, Cursor and JetBrains, in the Code tab of the desktop app, in the browser and in the Code tab of the Claude app on your phone. With Remote Control, you continue a session that runs on your computer from your phone or browser. On Team and Enterprise, an Owner has to enable it first. | Developers stay in their own environment. Anyone who doesn't want a terminal works in the desktop app. A long task keeps running while you're on the road. |
| GitHub Actions and automationno one at the keyboard | In GitHub Actions, Claude Code responds to @claude in a pull request or issue and can implement changes on its own. Team and Enterprise also get Code Review, where several agents read every change. From a script, Claude Code runs without an interface, through the Agent SDK. | Reviews and recurring checks run automatically with every change. |
What CLAUDE.md and skills look like at GO10X.
## Hard rules
- One CTA everywhere: the "Free AI Strategy Call".
- Brand name spelling: always `GO10X`.
- Deploy: NEVER run `wrangler deploy` by hand.
The only way to production is a PR
from `develop` onto `main`.# TODO
Open items for GO10X. Open items only.
What is done lives in the Git history.
## 2. Agent-ready
- [ ] AEO/GEO per the playbook
- [ ] ProblemStatement closer to the customer's language---
name: german-copy
description: GO10X-specific German copy rules,
locked claims, banned wordings, register
---
## Register
Informal "du" is locked for all marketing copy.- One CLAUDE.md per repository
- It holds the fixed rules: what must never happen, where the sources live, how things ship. Ours says, for example, that changes go live only through a pull request, never by hand.
- A TODO.md as the work list
- Every repository has a list that holds only open items. The instruction to the agent then reads: take the next item. Finished items get deleted, not checked off.
- Skills for everything that repeats
- A skill is a set of instructions with fixed inputs and a fixed result. Our language skills hold the approved wordings. Twelve skills live in a central repository and are linked into every project.
- One model plans, subagents do the work
- The main chat plans, delegates and reviews. Subagents with fresh context do the work, each with exactly one task. The strongest model judges, a cheaper one does the research.
- Checks before anything is published
- Before anything goes live, we run type checks, tests and an SEO check. Our HubSpot automation has more than 1,100 automated tests. The website fails the build when the headline on the preview card no longer matches the page.
- Connectors instead of retyping
- Through MCP, Claude reads directly from the CRM, inbox, calendar and browser. “Summarize this” becomes “file it, write the draft, log it in the CRM.”
We run it ourselves.
This website
We write, review and publish it with Claude Code. What gets decided in the morning is live the same day.
The Content Engine
Our own AI newsroom collects sources, rates them, writes the newsletter and sends it. We built it with Claude Code and run it ourselves. A new issue goes out every morning at 7 a.m.
HubSpot automation
Customer setup, invoices, contracts with e-signature and filing in SharePoint run without retyping, and a person approves. We built it with Claude Code, and it runs in three companies.
License Check for HubSpot
Our own tool reads a HubSpot portal and shows which paid features your team actually uses and which ones you don't need. We built it with Claude Code, and it has been in use since September 2026.
Four risks, four fixed rules.
Claude Code reads files, runs commands and acts in other systems. These are the four biggest risks in day-to-day use. For each one we follow a fixed rule, the same on every machine.
Our page on MCP servers shows how we secure individual connections.
Claude Code or Codex? And where Copilot and Cursor fit.
Claude Code and Codex have a lot in common. Both are AI agents that read code, edit files and run commands, in the terminal, in the code editor and in an app of their own. GitHub Copilot is built around GitHub and the code editor, and Cursor is a code editor itself. What each tool does is how the vendors describe it: Anthropic, OpenAI, GitHub and Anysphere, the company behind Cursor. When each tool fits is our assessment, as a firm that rolls out Claude Code.
| Tool | What it is | A good fit when | A weaker fit when |
|---|---|---|---|
| Claude CodeAnthropic | An agent harness from Anthropic, the software around the Claude models that turns them into a coding agent. Claude Code runs in the terminal, in VS Code, Cursor and JetBrains, in the desktop app, in the browser and on your phone. Skills and plugins also load in chat and in Cowork. | You want engineering and business teams to work with the same rules, skills and connectors, or your business teams already use Claude. | Your team wants to switch between models from different vendors in one tool. Claude Code works with Anthropic's models. |
| CodexOpenAI | OpenAI's AI coding agent. Codex runs in the ChatGPT app, in the terminal and as an extension for VS Code, Cursor, JetBrains and Xcode. In Codex Cloud, tasks keep running even while your computer is asleep. | Your company already uses ChatGPT Business or Enterprise. According to OpenAI, Codex is included in these plans, and you manage access and permissions in the same ChatGPT workspace. | Your business teams already work with Claude, and you want engineering to use the same models. In the ChatGPT workspace, Codex works with OpenAI's models. |
Alongside them: GitHub Copilot and Cursor
GitHub CopilotGitHub
GitHub's AI assistant, built into repositories, issues and pull requests. Copilot completes code as you type, reviews changes and takes on tasks from the issue to the pull request. This is not Microsoft 365 Copilot.
- A good fit when
- Your code lives on GitHub, and your team wants code completion in the code editor they already use, with models from OpenAI, Anthropic and Google to choose from.
- A weaker fit when
- You want sales or accounting to work with it too. GitHub describes Copilot as an assistant for software development.
CursorAnysphere
A standalone code editor with a built-in agent that imports settings and extensions from VS Code. Tasks go to Cloud Agents that work in parallel in their own virtual machines.
- A good fit when
- Your engineering team wants an agent, code completion and a free choice of model in one editor, from Anthropic, OpenAI and Google models to Cursor's own.
- A weaker fit when
- You want your team to stay in the code editor it already uses. Cursor is a separate application that you install next to VS Code or JetBrains.
Companies also use other tools, such as Google's Gemini CLI, which brings the Gemini models into the terminal, or Microsoft Copilot in Word, Excel and Outlook. This comparison stays with the four agents for software development.
- Why we roll out Claude Code
- We built our own operation with Claude Code: this website, the Content Engine, the HubSpot automation. The templates we bring come from there. Skills and connectors from one company repository also load in chat and in Cowork, so sales and accounting use the same workflows as engineering.
- Several tools in one repository
- All four tools read an AGENTS.md in the repository. If your rules live there, they apply to every tool. Claude Code reads the file in place of a CLAUDE.md or imports it from one, and it also runs as an extension in Cursor. Whichever of the four you pick, you need rules, central permissions and a review step. We use Claude Code, Codex and Cursor ourselves.
Common questions.
What is an agent harness?
An agent harness is the software around an AI model that turns it into a working agent. It calls the model again and again until the task is done, gives it tools such as reading and editing files or running commands, manages context and memory, sets permissions and limits, and checks the results. Anthropic itself calls the Claude Agent SDK, which provides the same agent loop and tools as Claude Code, an agent harness. OpenAI's Codex is a harness in the same sense. Setting up and extending this harness for a team is called harness engineering. For your company, this means you set rules, permissions and review steps in the harness. They then apply in every session, whichever model is running.
Is Claude Code GDPR-compliant, and where does our data go?
A company plan and central settings put the prerequisites in place. With personal accounts, it is out of your hands. Claude Code runs on your machines and sends prompts and responses, encrypted, to the model provider: Anthropic itself, Amazon Bedrock, Google Cloud or Microsoft Foundry. According to Anthropic's documentation, on Team, Enterprise, the API or one of these cloud providers, Anthropic does not train on your code. On Team, Enterprise and the API, Anthropic keeps data for 30 days by default. On Pro, Max and Ultra, the consumer plans, each user decides with a setting in their own account whether their data goes into training, and that includes your code. With Anthropic, the data processing addendum with EU Standard Contractual Clauses is part of the Commercial Terms. With a cloud provider, that provider's terms apply. Your IT uses managed settings to block files such as .env, turn off telemetry and approve connectors. Whether that is enough for your data is for your data protection officer to assess.
Which plan fits when a whole team works with Claude Code?
Anthropic recommends Claude for Teams or Claude for Enterprise for most organizations. The Team plan comes with SSO, admin tools and centrally managed settings for every Claude Code user. The Enterprise plan adds domain capture, role-based permissions and a Compliance API. If your infrastructure already runs on AWS, Google Cloud or Azure, Claude Code also works through Amazon Bedrock, Google Cloud or Microsoft Foundry, billed through your cloud account.
What does Claude Code cost, and how do we keep costs under control?
Claude Code is included in every paid Anthropic plan from Pro up. On Team and Enterprise, usage counts against each seat's allowance, which Claude Code shares with chat and Cowork. If you need more, you can turn on usage credits and set spend limits for the company, for groups or for individual employees. Current prices are on claude.com/pricing, because they change.
How do you roll out Claude Code for teams, and how long does it take?
Anthropic names five decisions for getting started: which provider you use Claude through, how settings reach your machines, what gets enforced, how you see usage and how data is handled. We make them with you and your IT. After that, the first workflow is live in one to two weeks. Then we move on, one use case at a time.
Where is your team stuck with Claude Code?
Bring what you've tried so far. In 30 free minutes you get a clear assessment of what's doable for you and what we'd advise against.
Prefer to read first? Subscribe to our AI briefing (in German)
