AI coding agents · Built in Switzerland
Everyone sells “autonomous”. Nobody sells “mergeable”. Anvil Coder takes a ticket from your backlog, breaks it into small work steps and delivers a draft pull request to your repository in which every step was produced under your rules and leaves its own commit. You read instead of reconstructing.
Describe what should change and how you will know it is done. Your repository and your project rules are the working context.
Anvil Coder plans the work steps, lets agents implement them in parallel and checks every step against the same guardrails. Every step becomes a commit.
The draft PR sits in your repository. You see what was asked, what changed and what was checked, and you decide on the merge.
Example
Thursday, 16:30. A small bug has sat in the backlog for three sprints: registration accepts an empty email address. Nobody had time. You write three sentences into the ticket and drag it to “Ready”:
A quarter of an hour later a draft PR with three commits is in your repository: fix, regression test, docs. You read the diff and the test report and merge if it fits.
How much oversight you keep is a per-project setting: from “every action needs an approval” to “only the consequential ones”. Without a set level a task runs through, and the result is in every case a draft PR that a human merges.
Your architecture and quality rules stand as guardrails in every single prompt. Whatever you store as a machine-readable rule is also checked during the run and blocks on violation.
An existing repository, understandable tickets and the build and test commands a person already runs today: that is all it takes. Java with Spring Boot is the most travelled path; other stacks run through the same stack profile as soon as their tools are in the verification environment.
Anvil Coder is developed and operated by Halvic Labs in Switzerland. You decide which model sees your code: Anthropic, OpenAI-compatible providers or your own key.