CODE68 v0.1.0
Setup Guide →
Code68 v0.1.0 Released · Cloud Intelligence + Local Execution →

C O D E 6 8

Autonomous AI Coding Agent for Your Terminal

Code68 pair programs with you directly in your terminal. Heavy cloud intelligence runs remotely on free GPUs, while surgical file edits, symbol searches, terminal commands, and automated test runs execute safely on your local machine.

Setup & Docs Guide Install Code68 Live Terminal Demo
> pip install git+https://github.com/PRN-6/code68.git

&zap; Run directly in any project folder. Global install, zero local GPU required.

0 Local GPU Required
33 Tests Passing
< 50ms Tool Latency
Python · Git · Shell · AST
Interactive Experience

See Code68 In Action

Watch how Code68 inspects code, applies surgical diffs, runs pytest, and maintains full context.

bash — code68 ~/my-project
Dual Tesla T4 Connected
Core Architecture

Cloud Intelligence + Local Execution

Your model never touches your filesystem directly. All operations pass through strict, local security gates.

Cloud Brain (Remote GPU)

Zero local RAM/VRAM load

Remote
  • ✓ Qwen2.5-Coder-7B (4-bit NF4 Quantization)
  • ✓ Free Kaggle dual NVIDIA Tesla T4 or Ollama
  • ✓ OpenAI-compatible SSE streaming API
  • ✓ Secure Bearer Token authentication

Local Agent (Developer PC)

Confined to current workspace

Local
  • ✓ Surgical unified diff file edits
  • ✓ Terminal command runner with timeouts
  • ✓ AST symbol indexer & ripgrep search
  • ✓ 5-Level security sandbox & audit logging
Engineered for Devs

Built Different. Truly Autonomous.

Surgical Diff Editing

Never rewrites your entire file blindly. Code68 generates unified git-style diffs, replacing only the exact target snippet with whitespace preservation.

Loop & Stall Detection

Autonomous agents can get stuck repeating actions. Code68 hashes tool invocations and immediately breaks repetitive loop cycles before wasting tokens.

5 Permission Levels

From strict read-only to edit-approved and full-auto. Dangerous actions like file deletion always prompt for confirmation.

Deep Codebase AST

Indexes function declarations, classes, and dependencies using Python Abstract Syntax Trees. Searches symbols across thousands of lines instantly.

Provider Agnostic

Swap freely between Kaggle, local Ollama, vLLM, OpenRouter, Groq, or DeepSeek without modifying a single line of client agent code.

Session Resume

Conversations are preserved in .code68/sessions. Use /save and /resume to pause and continue complex coding tasks anytime.

Quick Setup

Install Code68 in 30 Seconds

Available globally on Windows, macOS, and Linux. Requires Python 3.11+.

Looking for the complete deep-dive setup manual?

Detailed step-by-step guides for Kaggle dual Tesla T4s, Ollama, OpenRouter/Groq, security permissions, and CLI workflows.

Read Complete Docs →
Questions & Answers

Frequently Asked Questions

How does Code68 protect my private code and files? ↓

The cloud model never has direct access to your filesystem. All operations pass through explicit, sandboxed local Python tools. Code68 uses resolve_safe_path to prevent directory traversal attacks, masks API keys and secrets before logging, and blocks destructive commands like rm -rf.

Do I need an NVIDIA GPU on my local computer? ↓

No! That is the core architecture of Code68. The heavy neural network (Qwen2.5-Coder-7B) runs in the cloud (e.g. on Kaggle's free dual Tesla T4 GPUs or via OpenRouter/Groq), while your local machine only executes lightweight Python scripts and tools. You can run Code68 on an ordinary laptop.

Can I use Code68 on multiple projects? ↓

Yes! You configure your endpoint once using code68 setup (or code68 config --global), and then you can navigate into any directory on your computer (cd D:\projects\my-app) and type code68. It automatically binds to that project.

Can I run Code68 100% offline without the cloud? ↓

Yes! You can connect Code68 to your local Ollama instance (http://localhost:11434/v1). Run ollama run qwen2.5-coder:7b, and Code68 will execute 100% locally on your own hardware with no internet connection required.