You keep your process

Front: your ticket starts the run

The entry point is the same gesture on every tracker: you pull a story onto the start state, Anvil Coder picks it up, plans, builds and reports the outcome back into the ticket state. Pull anepic onto the start state and it decomposes into user stories that sit as real tickets in your tracker; which of them runs is your call, one by one. Close a ticket and the running work stops.

You can choose the internal issue board in the cockpit, YouTrack,GitHub Issues, Jira Cloud, GitLab Issues andBitbucket. Before the first run a board check verifies that your board knows every state the flow needs and proposes missing ones for confirmation. Planned: Trello and Linear.

From the pull request on you can stay there: a comment on the pull request becomes the requirement of an incremental follow-up run, and comments on the ticket in the review states trigger follow-up runs. Both channels are switchable per project and switched on in the rented service. The first run of a project always comes from the ticket. Triggers, permissions and the feedback channel per tracker are described in the guides (in German) tickets board andcreating a project.

Back: the result is a repository

Every completed work step leaves its own commit. What stands at the end is not a text block someone copies out, but a repository with history, readable for humans, verifiable for your pipeline. The result arrives where you review anyway: as a draft pull request against the main branch in your private repository. Nothing to import, the first look happens in the familiar review tool.

Available code hosts are GitHub, GitLab (gitlab.com today) andBitbucket Cloud; on GitLab and Bitbucket the counterpart is called a merge request, the flow is the same. Anvil Coder sets up the necessary webhooks per project itself. How to connect is described under connecting repositories (in German).

In the middle: the model is interchangeable

Anthropic

The path the system was built on, directly via the API or through the command-line integration.

OpenAI-compatible providers

One shared adapter serves OpenAI, Groq, Gemini, Mistral and Ollama. Which provider runs is decided by the deployment configuration.

Your key, your run

Each organisation stores its own model key and settles directly with the provider. On the Claude paths nothing runs without it; for OpenAI-compatible providers it overrides a configured default.

Local models

Ollama is addressed like any other provider. If you set a price per unit of compute time, runtime flows into the cost accounting; without a price, local models run at cost zero.

The brands behind it

Context from your systems

Three read-only context servers give the agent context the ticket itself does not carry: one for YouTrack stories, one for GitHub pull requests and one for a curated text corpus. You enable them as needed; by default they are off. An unassigned organisation gets nothing from the text corpus and from GitHub, not another organisation's data, nothing at all; for YouTrack the operator can configure a shared default instance that an organisation without one falls back to.