§ Guide · Claude Code for teams

    Claude Code is not autocomplete. It's a colleague that works in your terminal.

    Claude Code runs in your shell, reads your entire codebase, executes commands, writes pull requests and waits for review. Below: what it is, what it isn't, and how an Agentic Engineering team makes it production-ready.

    01 — Definition

    What is Claude Code?

    Claude Code is Anthropic's official CLI coding agent, launched in 2025 and built on the Claude model family (Sonnet, Opus, Haiku). Unlike an IDE plugin, Claude Code runs in your terminal: it has access to your entire working directory, can read and write files, run shell commands, run tests and perform git operations.

    It shifts the relationship with AI from 'autocomplete in my editor' to 'autonomous colleague I assign a task to and who comes back with a working PR'. That sounds small, but it changes how you structure your day.

    02 — Workflow

    How it works in practice

    You open a terminal in a project, start `claude`, and describe what you want done — 'add pagination to the orders endpoint and update the tests'. The agent plans the steps, reads relevant files, makes changes, runs the tests, and shows you the diff before committing.

    A `CLAUDE.md` at the root of your repo hands the agent your project conventions: coding standards, architectural principles, which commands to use for build/test/lint, and which directories are off limits. That file is the single source of truth for every agentic session.

    For repeatable tasks you build sub-agents (specialist roles like 'security-reviewer' or 'migrations-writer') and slash commands. The difference between amateur and professional use lives entirely in that configuration.

    Claude Code versus the alternatives

    DimensionClaude CodeCursorGitHub Copilot
    InterfaceTerminal / CLIIDE (VSCode fork)Plugin inside IDE
    ModelClaude Sonnet/Opus/HaikuMulti-model (incl. Claude)Mostly GPT family
    Scope per taskWhole repo, multiple filesWhole repo via editorMostly current file + context
    AutonomyHigh — plans, executes, testsMedium — propose + acceptLow — completes lines and blocks
    CustomisationCLAUDE.md, sub-agents, hooksRules, MCP, ComposerCustom instructions, Workspace
    Best forRefactors, migrations, multi-file featuresDaily editing + agent modeFast suggestions while typing

    03 — Application

    Workflows we build for clients with it

    • Automated dependency upgrades with test suite and changelog PR per package.
    • Codebase migrations (Vue 2 → Vue 3, REST → tRPC) as one agentic task per module.
    • Multi-model code review: Claude Code writes, a second agent (Codex or Gemini) reviews before the PR reaches a human.
    • Compliance checks: a sub-agent reads every PR and flags violations of internal policy or GDPR requirements.
    • Dev-on-call: a terminal agent that triages alerts, runs runbooks and only escalates when stuck.

    04 — Operations

    Governance, security and cost

    01

    Token and cost control

    Claude Code can burn hundreds of euros overnight. We set per-project budget limits, log every run, and route to Haiku or Sonnet depending on complexity (see our blog on tokenmaxxing).

    02

    Tool permissions and sandboxing

    By default the agent can run shell commands. We run it in a container with read-only volumes for sensitive paths and an allowlist of permitted commands.

    03

    No training on your code

    Anthropic's Enterprise contract excludes training on customer data. For those who want more: pair with OpenClaw for an on-prem inference layer so even metadata stays inside your network.

    04

    Auditability

    We persist every agentic session: prompt, plan, tool calls, diffs, outcome. Compliance and review become a query, not archaeology.

    05 — Risk

    Pitfalls teams fall into

    01

    No CLAUDE.md, no quality

    Without project instructions the agent falls back to generic patterns that don't match your codebase. CLAUDE.md is not optional.

    02

    Unbounded autonomy

    `--dangerously-skip-permissions` is named that way for a reason. In production: review gates and command allowlists always.

    03

    No evaluations

    If you don't measure progress, the agent doesn't make any. Eval suites against real tasks are mandatory.

    04

    Forgetting it's a colleague

    Treat the output like a junior engineer's: read, review, ask questions. Don't merge blindly.

    Our stance: Claude Code is an instrument, not a magic wand

    Claude Code is currently the strongest open-ended agentic coding tool we know. We use it daily — for our own code and, embedded in CI/CD, for client systems.

    The value isn't in the tool, it's in the engineering around it: clear instructions, evals, cost control, security sandboxing and human review at the right moments. That's Agentic Engineering.

    Claude Code production-ready in your team?

    We help engineering teams adopt Claude Code with the right guardrails: CLAUDE.md templates, sub-agents, eval pipelines, cost dashboards and — if desired — an on-prem variant via OpenClaw.

    Frequently asked questions

    Does Claude Code replace my developers?

    +

    No. It makes strong developers much more productive and weak developers visible. Scarcity shifts to people who can design, judge and orchestrate — Agentic Engineers.

    Can I run Claude Code on confidential code?

    +

    Under Anthropic's Enterprise contract your code is not used for training. For strictly confidential or regulated data we recommend a hybrid setup with an on-prem inference layer (OpenClaw) for the sensitive parts.

    What's the difference between Claude Code and Claude in the browser?

    +

    Browser Claude is a chat interface without repo access or tool execution. Claude Code runs in your terminal, reads your files, executes commands and delivers working diffs. Different mode, different capabilities.

    What does it cost in practice?

    +

    Highly dependent on model and task. A typical full-stack feature task on Sonnet costs a few tens of euros in tokens. An unmonitored agent in a loop can run up hundreds per day — which is why budget monitoring is mandatory.