Kite - Your Personal Git Assistant

Inspiration

Every leap in developer tooling started with a better way to interact with code. Git gave us history; hosts made collaboration routine; PRs made reviewable change the norm. But the interface hasn’t kept up with intelligence. Developers still juggle intent, context, and ceremony across commands, tabs, and tribal rules. We set out to make version control intent‑first and largely invisible by building a context engine that stays two steps ahead.

What it does

Kite is an AI Git assistant that lives in your terminal and bridges thought to action.

Natural language to precise Git: describe the task, get a previewed plan, then execute safely.
Conflict intelligence: a fine‑tuned model proposes explainable merge resolutions with side‑by‑side diffs and instant rollback.
Semantic commit splitting: turns sprawling edits into coherent, reviewable commits with clear messages and scopes.
Retrieval‑augmented repo chat: ask “why was auth refactored?” or “what changed since v1.3?” and get cited answers to commits/PRs/reviews.
Plan‑then‑act safety: dry‑runs by default, temp branches for every operation, one‑command undo.

Currently, Kite is just an agent with tool calls with a fine-tuned model and clever orchestration but our vision for Kite is it sitting on an emerging context engine: a memory layer over your codebase plus multi‑agent orchestration that predicts the next two steps based on your preferences and behavior.

How we built it

AI Agent Assistance
- Plans for branch strategy, semantic history, and CI/review gates.
- coordinates composes a safe, auditable trajectory (preview → execute → verify).
- ability for users to run the model locally my switching the provider via the CLI.
Conflict intelligence
- Fine‑tuned model whta we like to call KiteResolve20B on synthetic and real conflict corpora.
RAG for repo Q&A
- Hierarchical retrieval (file → hunk → commit/PR) with deterministic citations.
- Conversation state bound to repository snapshots.
CLI/Terminal UX
- Easy-to-use user friendly, CLI design aiming for a minimalistic and mode approach.
- Every Kite action creates a backup branch and lets you know the details of it, in case you ever want to go back, which you can do easily with Kite.

Tech stack:

TypeScript backend + Go (Charm library) for CLI
Python for training/inference
Next.js for Dashboard
GitHub APIs
ConvexDB for chat history and user settings;

Challenges we ran into

Building trust in automation: previews, diffs, and safe rollbacks were non‑negotiable.
Fine-Tuning the model:
Conflict datasets: curating realistic, diverse conflict scenarios that generalize.
Long‑context reasoning: staying faithful to repo state without blowing token limits.
Ergonomics in the terminal: Utilizing the right tech stack for keeping the flows fast.

Accomplishments that we're proud of

Plan‑then‑act flow that turns plain English into exact, previewed Git operations.
Conflict resolutions that are explainable and reversible, not black‑box merges.
Semantic commit splitting that consistently yields cleaner histories and faster reviews.
Repo Q&A with reliable citations back to commits/PRs, answers you can verify.
Fine-tuning the first ever comprehensive merge conflict resolver model powered by OpenAI GPT-OSS20B.
and our team :)

What we learned
Developers adopt AI that reduces workload without abstracting away guarantees.
Explanations + citations are as important as raw model capability for trust.
Ability for local‑first design meaningfully lowers adoption friction.

What's next for Kite - Your Personal Git Assistant

Live context engine in the terminal: preference and behavior‑aware prediction that proposes your next two steps before you ask.
Effectively “unlimited” context for the model: Our idea to solve this is by building "Infinite Chat" - a feature that gives your chat applications unlimited contextual memory. It works as a transparent proxy in front of your existing LLM provider, intelligently managing long conversations without requiring any changes to your application logic.
Cross‑tool memory: carry intent across editor, CLI, CI, and code review without new silos.
Policy‑aware automation: enforce conventions automatically, human‑readable history by default.