How Kelt Forge works

Point Kelt Forge at an existing repository and it reads the files a request actually touches — not just the one file you had open — and works from your project's existing structure and conventions. Or describe a new idea in a sentence and Kelt Forge starts a new project from scratch, with an auto-detected stack or one you choose. Both paths lead into the same workflow below.

Before a single line changes, Forge AI produces an ordered plan: the files it will touch, what it will add, what it will delete, and what it is unsure about. You see the plan before any edit happens.

Forge AI can read a task and choose the right model for it automatically — weighing capability, speed, cost, availability, and the task at hand — across model families including Claude, Grok, OpenAI models, DeepSeek, and Kimi. You can also pick a model yourself. Either way, the work runs through the same reviewed, tested, Git-backed workflow.

Forge AI makes the edits directly in your project, then runs your own test suite — not a synthetic one — and reads the failures, iterating where it can or stopping and reporting exactly where it got stuck. An agent that writes code without running it is a suggestion engine; Kelt Forge runs what it writes.

For projects with a running app, Kelt Forge maintains a live preview of the current state alongside the code and chat — real project files behind a real deployment, not a mockup.

The output of every task is a real Git commit in your own project, with a summary of what changed and why — reviewable diff, real file paths, nothing hidden behind the chat window.

Nothing is force-pushed and nothing merges to your main branch without your approval. Every change is a normal commit, reviewed and reverted the way you'd handle any other commit.

Alongside your code, Kelt Forge keeps a project brain — a set of plain-markdown files (architecture, decisions, open todos, known bugs, completed features, conventions, and more) that Forge AI reads and updates as it works, so context about your project persists across sessions instead of resetting with every new chat.

Kelt Forge is built like infrastructure around a real repository and a real test suite, not a chat window that returns a code snippet. It reads whole-project context before editing, runs your tests before calling anything done, and its deliverable is a Git commit your team already knows how to review — not a copy-pasted answer.