Fragon Forge is our own product: an AI agent that turns a task description into a finished proposed change to the code, known in the trade as a merge request. We use it in our own projects, and the commercial launch is planned for the first quarter of 2027. For you as a client, this text is mainly evidence of how we build software that many customers use at the same time, and safely.
At a glance
| Project | In-house development, Fragon Studios |
| Industry | Tools for software developers |
| Technologies | .NET 10, React 19, TypeScript, PostgreSQL, Qdrant, Kubernetes, MCP, OpenCode |
| Period | since 2025, ongoing |
| Status | In use in-house, commercial launch in the first quarter of 2027 |
| Result | From task to merge request without manual work, for four Git platforms |
The starting point
A large part of software development is repetitive work: writing boilerplate, updating libraries, reviewing code, looking for security holes. As a small software studio, we asked ourselves whether an AI agent could do this work on its own, within the usual Git workflow.
The available solutions (GitHub Copilot, CodeRabbit, Cursor and others) were tied to GitHub, covered only part of the work or required a workflow that had little to do with the team's own Git platform. None of them suited teams on GitLab, Gitea or Bitbucket with strict GDPR requirements.
We first built Fragon Forge as a tool for our own repositories. After a few months of using it ourselves, it was clear that many other teams were missing the same thing. That is how the internal tool became a product.
The challenge
Turning an internal tool into a product means solving several problems at once.
Four Git platforms, one workflow
GitHub, GitLab (CE and EE), Gitea/Forgejo and Bitbucket Cloud each have their own APIs, their own formats for notifications (webhooks) and their own authentication methods. The shared layer on top of them has to be clean, otherwise the effort grows with every platform.
Writing and reviewing in one run
A developer describes a task (an issue), the AI delivers a merge request and then reviews it itself for security problems, vulnerable libraries and style.
Running other people's code safely
An agent that writes code on behalf of a customer also runs that customer's code: tests, builds, analysis tools. In a free plan, that means running any code from anyone, many times a day. Without strict isolation, that is a security risk.
Control over your own data
Operation in the EU, optionally with your own key for the language model (BYOK, Bring Your Own Key), and no training on customer code. Anyone who uses Fragon Forge should know where their data is stored and who sees it.
Our solution: Fragon Forge
Fragon Forge connects to your Git platform via webhook and has four core functions.
1. Coding agent: from task to merge request
A developer creates an issue and adds the label ai-task (or calls the agent with @fragonforge in a comment). Fragon Forge receives the webhook and clones the repository, and an AI agent reads the task and writes the code. Fragon Forge then opens a draft merge request together with its own review. The developer checks it and accepts it.
What this means for your project: the key decision was not to build the agent as a chat. It gets the full context once and delivers a result instead of asking follow-up questions. The developer does not have to wait and can do something else in the meantime.
2. Static analysis: more than ten tools, one dashboard
An analysis runs with every merge request: Semgrep and Trivy for security and container vulnerabilities, Gitleaks for credentials checked in by mistake, Checkov and trivy config for infrastructure code (Terraform, Kubernetes, Helm, Bicep), plus ESLint, RuboCop, PHPStan, Ruff and Clippy for individual languages, and Lizard and jscpd for complexity and duplicate code.
The results go into a dashboard with a grade from A to F, a breakdown by severity and a history over time. Only new findings block the merge request. Existing legacy issues do not hold up the work. If the agent finds something it can fix itself, the label autofix-ready triggers a fix run.
The tools run inside the same isolated run as the coding agent, not as a separate CI pipeline. As a result, the findings are stored in structured form in the database, and the dashboard can show trends across all branches and repositories.
3. Several agents with separate roles
A single agent works up to a certain task size. Beyond that, it gets in its own way. Fragon Forge therefore separates planning, implementation and review, each with its own language model. Stronger models suit planning (Claude Opus, for example), faster ones suit implementation (Claude Sonnet, for example).
In specification mode the chain becomes explicit: the issue first becomes a specification, then tests, then code, with a human approving in between. For whole projects there is detailed planning with exploration, architecture, a breakdown into subtasks and a final assessment.
4. Knowledge of the repository and lessons learned
The agent reads your repository with tree-sitter and stores it in a vector store (Qdrant). That way it finds relevant code across the whole project, not just in the file it is working on. What it has learned from earlier tasks is saved, both in general and per repository. Comments on a merge request from Fragon Forge restart the agent with the full context, so feedback does not get lost.
What we dropped: one container per customer
The first version looked different. Every customer got their own Fragon Forge container with its own database and its own storage. That seemed the obvious choice, because physical separation is the simplest picture of isolation: no shared rows, no forgotten WHERE clauses, no shared queue.
It was clean, too. But it does not scale.
We had underestimated three things.
The cost of a free plan
One container per customer means one container per account with a fixed baseline consumption, whether the account is active or not. In a free plan without a credit card, that turns into hosting costs that make the plan impossible.
The operational effort
Hundreds of containers mean every update, every monitoring check, every backup and every database change a hundred times over. On top of that come a reverse proxy with automatic configuration, authentication between the control plane and the customer containers, and different versions per container. What is simple when done centrally turns into coordination work.
The wrong kind of isolation
Separate data per container answers the question of whether another account can see your rows. But the real risk with a coding agent is that it runs other people's code. One container per customer does not help against that, because that container runs the customer's code anyway. The isolation question sits one level deeper than the tenant question.
We made the change step by step over several versions rather than as a rebuild. That was slower, but every intermediate state kept working.
Today there is one shared PostgreSQL database with filters per tenant (data separation), a short-lived Kubernetes job in its own virtual machine for each run (execution separation) and an MCP gateway as the only way out of the sandbox (permission separation). Those are three layers for three different risks, and one container per customer would have solved none of them.
How we run code safely
At Fragon Forge, many accounts share one application and one database. Data and runs are kept apart by the architecture, not by trust.
Data separation
One PostgreSQL database, one schema, and every row of a tenant carries that tenant's ID. The separation is enforced in the data access layer (EF Core Global Query Filters), on every request and in every background job, never through individual WHERE clauses that someone could forget. A test with two tenants checks every endpoint, every background path and every search query for leaks. No version ships without passing this test.
Execution separation
Letting the agent loose on someone else's repository means running arbitrary code. Every run is a one-off Kubernetes job: no root privileges, a read-only file system, limits on CPU time, memory, processes, disk and duration, and outbound network access blocked. In every plan, each run gets its own virtual machine (Kata Containers). The machines for agent runs are kept separate from those for the web, the database and the control plane.
No credentials in the sandbox
The agent knows neither the Git credentials nor the key for the language model. The trusted side runs git clone and git push, while the sandbox gets a working copy without credentials and hands the changes back. A proxy inserts the language model key separately for each run. The only way out of the sandbox is through the MCP gateway (Model Context Protocol) and this proxy.
Architecture decisions and what we learned from them
One interface for four Git platforms
Instead of building a separate path through the code for each platform, there is a single interface (IForgeClient) that holds the platform specifics: webhooks, merge requests, comments on lines of code, authentication. That keeps four integrations maintainable.
Each role gets its own model
Several agents only work with clearly separated roles. Planning may take longer and cost more, implementation has to deliver quickly and consistently, and review needs knowledge of the tools. One model for everything is always a compromise.
Your own key or an included budget
Language model prices fluctuate, and many customers already have a contract with a provider. So every customer can store their own key, and the provider then bills the customer directly. Customers who do not want their own contract use the AI budget included in the paid plans.
Settings on three levels
Settings apply in a fixed order: for the whole platform, per repository in the dashboard, and in the repository itself (a .fragonforge.yml file overrides everything). This lets teams manage their own settings without needing admin access.
Status and pricing
Fragon Forge supports four Git platforms (GitHub, GitLab, Gitea/Forgejo, Bitbucket Cloud) and brings more than ten analysis tools together in one dashboard. Separate roles for several agents, search across the whole repository, saved lessons and a dedicated virtual machine per run are part of the scope. We use it in our own repositories.
The commercial launch is planned for the first quarter of 2027. Prices according to fragonforge.com, as of November 2026:
| Plan | Price | Includes |
|---|---|---|
| Free | €0 | 1 repository, 1 user, 100 runs a month, with your own key |
| Solo | €29 a month | 3 repositories, 1 user, 300 runs a month, €7.50 AI budget |
| Team | €99 a month | 10 repositories, 10 users, 1,500 runs a month, €25 AI budget |
| Business | €299 a month | 30 repositories, 30 users, 5,000 runs a month, €75 AI budget |
Fragon Forge is operated in the EU, on self-managed Kubernetes at Hetzner in Germany.
What you can take away for your projects
AI agents work best without follow-up questions
A task with the full context delivers better and faster results than a chat that asks follow-up questions.
Separate the roles, for agents too
One agent for everything gets worse as the task grows. Clear roles, each with its own model, hold up longer.
The obvious isolation is often the wrong one
Physical separation is reassuring, but it only addresses one risk. If you run other people's code, you need separation on at least three layers: data, execution and permissions.
Put settings where the team works
Settings for the platform, per repository and in the repository itself scale better than a central settings page, because teams keep responsibility for their own settings.
Why we are sharing this
Fragon Forge teaches us how to build software for many customers: keep data separate, isolate other people's code and abandon a wrong approach in time. We bring the same way of working to client projects.
Frequently asked questions
What is Fragon Forge?
Fragon Forge is an AI agent that turns a task description into a finished proposed change to the code (a merge request) and then reviews it itself. It is a product of Fragon Studios from Linz. We use it in our own projects, and the commercial launch is planned for the first quarter of 2027.
Which Git platforms does Fragon Forge support?
GitHub, GitLab (CE and EE), Gitea/Forgejo and Bitbucket Cloud. It connects via a webhook, and the workflow is the same on all four platforms.
How much does Fragon Forge cost?
There are four plans: Free for €0, Solo for €29, Team for €99 and Business for €299 a month. These are the prices on fragonforge.com as of November 2026. The free plan works with your own key for the language model, and the paid plans include an AI budget.
Where are code and data processed?
In the EU, on self-managed Kubernetes at Hetzner in Germany. No model is trained on customer code. If you wish, you can store your own key for the language model, and the provider then bills you directly.
How does Fragon Forge stop other people's code from causing damage?
Every run is a one-off Kubernetes job in its own virtual machine (Kata Containers), with no root privileges, a read-only file system and outbound network access blocked. The agent knows neither the Git credentials nor the key for the language model. The only way out of the sandbox is through a gateway and a proxy.
Fragon Forge is a product of Fragon Studios e.U. Commercial launch: first quarter of 2027. More at fragonforge.com.
