AI coding extensions will make you a 10x developer…

Key Concepts

• AI Coding Extensions: Third-party tools integrated into IDEs (like VS Code) that leverage AI for coding assistance.
• Open Router: A unified API gateway that allows users to access various AI models (OpenAI, Anthropic, Gemini, Mistral, Deepseek, Qwen, etc.), both open-source and closed-source, through a single API key, often with better pricing and uptime.
• Client: A beginner-friendly, open-source AI coding agent extension, serving as the foundational tool in the discussed lineage.
• Rue Code: A fork of Client, offering more advanced features, customizability, and control, aimed at intermediate users.
• Kilo Code: A fork of Rue Code, combining the best features of Client and Rue Code with a cleaner UI and rapid development, currently the speaker's favorite.
• Codebase Indexing: The process of converting a codebase into high-dimensional vector embeddings to enable semantic search, allowing AI tools to quickly find relevant files based on meaning rather than just keywords.
• Context Window: The maximum amount of text (tokens) an AI model can process at one time, crucial for understanding the scope of a coding task.
• Tokens: The basic units of text that AI models process, directly impacting API costs.
• Reasoning Effort: A setting in some AI coding tools (like Rue Code and Codex) that controls how much time and computational resources the AI spends on planning and problem-solving before generating code.
• Vibe Coding: A term used to describe a superficial approach to AI coding, often involving single prompts for quick, low-effort results, contrasted with "AI Programming."
• AI Programming: A more serious, architectural approach to building software with AI, involving planning, design, and structured development.
• GLM 4.6: An open-source coding AI model highlighted as currently the best in its category, outperforming some closed-source models on specific benchmarks.
• Codex (OpenAI): OpenAI's comprehensive AI coding ecosystem, including an IDE extension, cloud-based agents, and GitHub PR review, powered by models like GPT-5 Codex.
• Claude Code (Anthropic): An AI coding agent known for its polished terminal UI, but limited to Anthropic's Claude models.
• MCPS (Multi-Agent Communication Protocols): A concept for connecting different AI agents or tools to work together, e.g., Vectal MCP.
• Git/GitHub: Git is a version control system for tracking changes in code, and GitHub is a platform for hosting Git repositories, essential for professional software development.

---

The Rise of Next-Generation AI Coding Extensions

The video highlights a significant shift in AI coding, moving beyond traditional tools like Cursor and Claude Code to a new generation of AI extensions. David Andre, founder of vector.ai, explains that these extensions are gaining traction due to several key advantages:

1. Ubiquity: They can be used in any preferred code editor, such as VS Code, eliminating the need to switch to a specific environment like Cursor.
2. Provider Agnosticism: These modern extensions support a wide range of AI model providers (OpenAI, Anthropic, Gemini, Mistral, Deepseek, Qwen, etc.), unlike Claude Code which is limited to Anthropic models. The video specifically recommends Open Router (openrouter.ai) as a solution to access virtually any model, open-source or closed-source, through a single API key, often with better uptime and cost efficiency.
3. Cost-Effectiveness & Open Source: Many of these extensions, like Kilo Code and Client, are free and open-source, eliminating the need for expensive subscriptions (e.g., $200/month for Claude Code). Their open-source nature allows users to inspect the code, contribute improvements, or fork projects.
4. Customizability & Transparency: They offer superior context engineering for power users and provide transparent usage overviews, including token counts and API costs. This contrasts with tools like Cursor, which have been criticized for a lack of transparency regarding usage limits.

Understanding Client, Rue Code, and Kilo Code

The video details the relationship and distinct purposes of three prominent open-source AI coding extensions:

Client: The Foundation for Beginners

• Origin: Client was created as an open-source coding AI agent.
• Target Audience: Most beginner-friendly, with a simple setup and less customizability.
• Setup Process:
  1. Install the "Client" extension in VS Code.
  2. Open the primary sidebar to access Client.
  3. Go to settings -> API configuration.
  4. Choose Open Router as the provider.
  5. Create an account on openrouter.ai, add credits (e.g., $5), and generate an API key.
  6. Paste the API key into Client's settings.
  7. Select a model (e.g., GPT-5, GLM 4.6), enable "thinking," and set token limits. Client automatically loads correct pricing.
• Key Features:
  • Act/Plan Mode: Allows users to choose if Client should make changes or just plan.
  • Advanced Context Usage: Provides a transparent overview of context length, input/output tokens, API cost, and size in kilobytes – a feature often missing in proprietary tools.
• Philosophy: Client intentionally avoids codebase indexing to give users full control over which context is inserted into the prompt, potentially saving API costs.

Rue Code: Client with Power-Ups for Intermediate Users

• Origin: R Code was forked from Client.
• Target Audience: Best for intermediate coders who desire more control over system prompts, models, and customizability than Client offers.
• Key Features:
  • Slash Commands: Similar to Claude Code, allows users to define custom commands for frequently used prompts, saving time.
  • Multiple Specialized Modes: Includes "Architect," "Code," "Ask," "Debug," and "Orchestrator" modes, which are pre-built prompts tailored for different parts of the codebase or development tasks. A "Mode Marketplace" allows browsing community-created modes.
  • Codebase Indexing: This is the biggest difference from Client. Rue Code indexes the entire codebase, converting it into a multi-dimensional vector database for semantic search.
    • Benefit: Enables faster, more relevant search (e.g., searching "authentication" retrieves all related files, even if the word isn't explicitly mentioned).
    • Controversy: While beneficial for large codebases, it can load unwanted context and potentially increase API costs. Rue Code allows users to disable indexing if preferred.
  • Memory Bank Support: Helps maintain project context and coding standards across sessions.
  • Enhanced Context Understanding: Better handling of large codebases, likely due to indexing.
  • Reasoning Effort: Allows users to select the AI's reasoning intensity (e.g., "high") for more thorough planning.
  • MCPs (Multi-Agent Communication Protocols): Supports connecting different MCPS, including Vectal's MCP, to manage tasks directly from Rue Code.

Kilo Code: The Best of Both Worlds with a Clean UI

• Origin: Kilo Code was forked from Rue Code.
• Target Audience: Aims to combine the best parts of Client (simplicity) and Rue Code (features) with a cleaner, less cluttered UI, making it the speaker's current favorite.
• Key Features:
  • Cleanest UI & Most Customizability: Offers a streamlined user experience while providing advanced controls.
  • Autocomplete: Features like "pause to complete" (triggers autocomplete after a set delay) and "manual autocomplete" (guesses intent for quick fixes/refactors).
  • Quick Task (Command K/I): Similar to Cursor's Command K, allows targeting isolated code sections for specific actions.
  • Rapid Development: The Kilo team is noted for its fast pace in adding new features and models (e.g., GLM 4.6 was added faster than Client).
  • Architect Mode Demonstration: The video showcases Kilo's "Architect" mode by building an MVP for "YouTube Jobs for AI Video Creators."
    • The agent asks clarifying questions about deployment and workflow to scope the project.
    • It creates a comprehensive to-do list and markdown files (e.g., architecture.md) for planning before writing any code.
    • This process exemplifies "AI Programming" – a structured, design-first approach, contrasting with "Vibe Coding" (single-prompt, low-level).
    • The agent can run for extended periods (10-20 minutes) to build a proper codebase structure, demonstrating its capability for serious projects.
  • Visual Context Overview: Kilo provides a beautiful visualization of the context window, allowing users to see what's taking up tokens and click into AI responses.
  • Checkpoints: Offers internal checkpoints, though the speaker emphasizes the importance of learning and using Git for professional version control.
  • Team Plan: Kilo offers a team plan with features like:
    • Smart Controls: Prevents data leakage by restricting models that train on user data.
    • Zero Lock-in: Allows using Kilo's API, own keys, or local models.
    • Cost Savings: Helps identify high-performing team members and ensures everyone uses the latest, most efficient AI models.

Deep Dive into GLM 4.6 Model

The video highlights GLM 4.6 as a groundbreaking open-source coding model that many people are unaware of.

• Performance: Benchmarks show GLM 4.6 outperforming Claude Sonnet 4.5 (Anthropic's latest) on several key metrics:
  • AME25 (Math): GLM 4.6 is the best.
  • Life Codebench v6: Absolutely destroys Sonnet 4.5.
  • HLE (Tool Calling): Amazing performance.
  • Browse Comp: Also beats Anthropic.
  • It nearly matches Sonnet 4.5 on GPQ8 (Google Proof Question Answer) and is the best open-source model on SWB Bench Verified and TS Squared Bench.
• Cost-Effectiveness: GLM 4.6 costs approximately 1/7th the price of Claude Sonnet 4.5.
  • Sonnet 4.5: $3 per million input tokens, $15 per million output tokens.
  • GLM 4.6: $0.5 per million input tokens, $1.75 per million output tokens.
• Trend: An analysis by Client's creators shows that open-source models (like Llama 2, Mistral Large, Deep Seek R1, Qwen 3, and now GLM 4.6) are rapidly closing the performance gap with closed-source models. It's predicted that within 3-6 months, open-source models could dominate.

Comparison with Claude Code and Codex

The speaker acknowledges that while the new extensions are powerful, established tools like Claude Code and Codex still have their place:

• Claude Code:
  • Strengths: Pioneered AI coding agents, offers a highly polished UI in the terminal.
  • Weaknesses: Limited to Anthropic models, which restricts access to other cutting-edge models like GLM 4.6. The speaker notes using it less frequently now.
• Codex (OpenAI):
  • Strengths: Currently considered the most powerful AI coding tool due to the GPT-5 Codex model (available via Open Router).
    • GPT-5 Codex: Described as a "beast" on medium and high reasoning efforts, highly reliable, and capable of adapting its planning time (30 seconds for simple changes, 10-15 minutes for risky refactors/new features).
    • Ecosystem: Offers a comprehensive suite including a cloud-based agent (for asynchronous tasks), a GitHub PR review feature, and seamless integration between its web agent and IDE extension for managing tasks.
  • Weaknesses: Limited to OpenAI models.
  • OpenAI's Aggression: OpenAI is rapidly developing Codex, aiming to surpass competitors like Anthropic and Cursor.

Conclusion: Adaptability is Key

The AI coding landscape is evolving rapidly, with new models and tools emerging weekly.

• While Codex currently holds the "king" status due to GPT-5 Codex, the speaker emphasizes that this could change quickly with the release of models like Gemini 3 or new versions of GLM/Deepseek.
• The key takeaway is the critical importance of adaptability. Developers must be willing to experiment with new tools and models, continuously learn, and master their setup to stay on the cutting edge.
• The new generation of AI coding extensions (Client, Rue, Kilo) offers unparalleled flexibility, cost-effectiveness, and access to powerful open-source models like GLM 4.6, making them indispensable for modern AI programming.