Claude Code in Frankfurt: the coding agent and what a bank settles first

According to Anthropic's product page, Claude Code maps and explains a codebase, reads issues and writes code, runs tests, opens pull requests and follows imports across repositories, and it keeps working when something breaks. You steer it from a terminal, from VS Code and Cursor, from JetBrains IDEs, from Slack, from the web, from the Claude desktop app, from iOS and Android, and from GitHub Actions, on macOS, Linux and Windows. It asks for permission before it changes a file or runs a command.

The two confusions worth clearing up: it is a different product from the Claude chat apps, where you paste text and get text back, and it is a different thing from an autocomplete extension, which predicts the next few lines without understanding the task. Claude Code takes a task, not a line. On Team and Enterprise plans the Projects feature supervises several agents at once, so one person can have three tasks in flight.

Three jobs come up again and again in a bank or an asset manager. The first is understanding an unfamiliar service: a payment adapter written six years ago by someone who left, with no documentation and a test suite that half runs. The agent reads it, follows the imports and explains what the thing actually does, which is a week of a developer's time.

The second is tests. Legacy finance code is often under-tested exactly where it matters, in the rounding, the currency conversion and the cutoff handling, and writing those tests is work nobody volunteers for. The third is migrations: a library version, a framework upgrade, a change of date handling across 200 files. That work is mechanical, large and error-prone by hand, which is the profile the agent suits.

What it does not take over is the decision. Somebody still reviews the pull request, and in a regulated firm that review is the control the auditor will ask about.

A repository can carry its own instructions file, which is where a team writes down the conventions, the commands and the rules for what the agent may touch. Skills package a procedure the team repeats so the agent follows the house method instead of inventing one. Subagents split a large task so one part of the work runs separately from the main thread. Hooks run a command of yours at defined points, which is how a team enforces a check the agent cannot skip.

For a finance team those four are the control surface: the instructions file sets the boundary, the hooks enforce it, and both live in version control where a reviewer can see who changed them.

OpenAI Codex is the comparable product from OpenAI, with a CLI, an IDE extension, a cloud service and the ChatGPT apps as its surfaces, plus sandboxing, agent approvals, auto-review and profile-based permissions. The two tools are built on the same idea and differ in the surfaces they offer and the platforms they assume.

Neither is ranked above the other here. ChatGPT and Claude compared for finance goes through the comparison, including the coding one, and explains why a platform team usually decides on the cloud and the identity provider it already runs.

Five things. Where the code goes: the tool runs locally and talks to the model API without a backend server of its own, and Anthropic offers self-hosted environments that run inside a customer network next to internal services, which is the answer to the hardest question in this list.

Then: which repositories it may touch and under which account, whether it may run commands that reach the network, how its actions appear in the audit log, and who reviews its pull requests. The permission prompt is the mechanism, but the policy has to exist on paper first, because a developer who clicks through every prompt has disabled the control.

The EU AI Act bears on the output and not on the tool: code that ends up in a system touching credit or customers lands in the obligations attached to that system, whoever wrote it.

At the German-language Claude meetup that the Agentic AI Community Frankfurt runs every other week at byte5 GmbH in the Speicherstraße, where the subjects include programming and web development alongside prompting and AI workflows. Claude meetups in Frankfurt describes the format and the room.

Finance Loop runs Claude Hacker House for people who work in finance, and the CCAF certification is the route for someone who wants the architecture side on paper. Blockchain software development covers the other developer field this site follows.