Yapie0

Yapie0 / carboncode

Public

Chinese-first DeepSeek-powered terminal coding agent

yapie0.github.iocarboncode ai-agent carboncode cli coding-agent deepseek
24
2
100% credibility
Found May 26, 2026 at 24 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
TypeScript
AI Summary

Carbon Code is an AI-powered coding assistant that runs in your terminal. It reads your code files, understands your project structure, and helps you write code, fix bugs, or understand how existing features work. You type questions in plain English, and the assistant reads your files to give accurate answers. When it wants to make changes, it shows you exactly what will change before touching anything—so you stay in control. It can run shell commands, search across your codebase, and manage multiple work sessions. The tool also includes a web dashboard to monitor your usage, costs, and past conversations.

How It Works

1
💻 You install the assistant

You download and install Carbon Code on your computer using a simple command, just like installing any other development tool.

2
🔑 You connect your AI account

On first launch, a friendly setup guide helps you connect your DeepSeek account so the assistant can think and help with your code.

3
📁 You open your project

You navigate to any folder on your computer containing code and tell the assistant to start working there.

4
🤖 You ask the assistant anything

You type questions in plain English: 'How does this feature work?' or 'Can you add user authentication here?' The assistant reads your files and thinks through your codebase.

5
The assistant takes action
💬
Explaining code

The assistant reads your files and explains how something works, points out bugs, or suggests improvements.

✏️
Writing or editing code

The assistant shows you exactly what it wants to change, line by line, and waits for your approval before touching anything.

6
✅ You review and approve

You see a clear side-by-side view of old vs new code. You can approve all changes, reject specific ones, or ask for refinements.

🎉 Your project is improved

With your approval, the changes are applied to your files. Your code is cleaner, new features are added, and your work is saved in a session you can resume anytime.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 24 to 24 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is carboncode?

Carbon Code is a terminal-based coding agent that lives in your CLI. You cd into any repository, invoke the agent, and it reads code, proposes edits, asks for confirmation before running shell commands, and maintains a session trail. Built entirely in TypeScript with a web dashboard for monitoring usage and budgets, it treats DeepSeek as its sole backend and ships with a full toolkit of built-in tools for file operations, web search, subprocess management, semantic search, and extensible MCP server integration.

Why is it gaining traction?

The hook is cost control through prefix-cache optimization. Rather than treating every API request fresh, Carbon Code freezes the system prompt and tool schemas at session start so DeepSeek's cache key stays identical across turns. The benchmarks show real users hitting 99.8% cache hit rates and cutting API spend by roughly 80% compared to naive implementations. It also wins on UX: edits require a review step before applying, shell commands are gated behind an allowlist, and the dashboard gives you live stats on tokens, cache hits, and budget burn.

Who should use this?

Individual developers who work in the terminal and want an AI assistant with real file system access but without the risk of an agent running wild. Teams evaluating DeepSeek for internal tooling will find the cache economics compelling. Developers who live in China or work with Chinese-language codebases get first-class localization, though the English interface is fully supported.

Verdict

With 24 stars and a 1.0% credibility score, this is an early-stage project that shows genuine engineering thought around cost efficiency and safe autonomy. The cache-first design is technically sound, the permission model is thoughtful, and the dashboard adds useful observability. Try it on a personal project, run the built-in benchmarks against your own workload, and measure whether the cost savings justify the dependency on a specific model provider. The risk is maturity: test coverage, documentation, and community support are thin, so treat it accordingly.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.