Andy Devdan discusses a strategy to optimize AI coding by leveraging new open-source models and the Model Context Protocol (MCP) to reduce costs while maintaining high intelligence. He explores the benefits of delegating AI coding tasks within Claude Code to other models using tools like Ader.
The AI Coding Dilemma: Time vs. Money
Engineers face a constant trade-off: spend money to save time or spend time to save money. While time is invaluable, the emergence of powerful, low-cost models offers a way to achieve both.
Meta's Llama 4 and the Rise of Open-Sourceish Models
Meta's Llama 4 family (Herd, Behemoth, Maverick, and Scout) are promising open-source models. Maverick, in particular, is competing with closed-source LLMs like ChatGPT-4 and Gemini 2.5 Pro. This competition signifies a shift towards high-quality open-source options, offering more choices for compute.
Claude Code and the Agentic Coding Revolution
Claude Code, powered by Claude 3.7 Sonnet, is a leading agentic coding tool that is transforming software engineering. It combines agentic coding, customizable MCP servers, and a powerful LLM.
The Costly Reality of Claude 3.7 Sonnet
A major drawback of Claude 3.7 Sonnet is its high cost, specifically the $15 per million output token price. This cost can be significant when writing code at scale, particularly when packaging and handing off large tasks.
The Solution: Delegating AI Coding with MCP and Ader
The solution involves delegating AI coding tasks away from Claude Code to other models, such as Behemoth, Maverick, Scout, or Gemini 2.5 Pro using the Model Context Protocol. This can save money and grant more control over the chosen model.
Advantages of Delegating AI Coding
-
Cost Savings: Utilize cheaper models for AI coding tasks.
-
Model Control: Regain full control over the model used within Claude Code.
-
Agentic Pattern: Tap into a powerful agentic pattern for improved workflow.
Demonstrating Ader within Claude Code: A Practical Example
Andy demonstrates using Ader within Claude Code to replace .info calls with .debug calls, showcasing the benefits of delegated AI coding.
Initial Setup and Context Priming
The process begins with running a context priming prompt to inform Claude Code about the codebase and instruct it to use Ader for coding tasks. This effectively overrides Claude Code's default file editing tools.
Selecting a Model and Executing the Task
The demonstration utilizes Gemini 2.5 Pro as the chosen model. By using Ader, the AI coding process is delegated to Gemini 2.5 Pro, resulting in a cost-effective solution without sacrificing intelligence.
Scaling Impact and Showcasing Token Savings
By delegating the AI coding process, significant output token savings are realized. Cloud Code acts as a tool orchestrator, while Ader and the chosen model handle the actual code writing.
Listing Models with Ader
Ader allows users to list available models from various providers, enabling selection of the most suitable model for a given task.
The Looping Editor Reviewer Workflow
Claude Code validates the changes made by Ader, creating an editor-reviewer workflow. This process ensures accuracy and quality, demonstrating the power of combining multiple tools and agents.
The Power of Combined Workflows
Andy stresses the importance of thinking in "ands" rather than "ors," highlighting the benefits of combining multiple tools and ideas for optimal results. The collaboration between Claude Code and Ader exemplifies this concept.
Cost Analysis and Leaderboard Insights
The demonstration concludes with a cost analysis showing significant savings due to reduced output tokens. Additionally, the Ader leaderboards reveal Gemini 2.5 Pro's impressive performance, surpassing Claude 3.7 in certain areas.
Key Takeaways and Future Predictions
-
MCP servers offer a way to delegate work and reduce costs.
-
Combining MCP servers with AI coding tools provides significant benefits.
-
AI coding models are constantly evolving, with new options emerging.
-
OpenAI is expected to maintain its lead in raw intelligence with upcoming models.
Ader: A Programmable and Customizable AI Coding Tool
Ader stands out as a programmable AI coding tool that can be integrated into various workflows and MCP servers. Its customizability and control make it a valuable asset for principled AI coding.
The State of AI Coding and High Leverage Bets
Andy encourages viewers to explore his "State of AI Coding" essay for insights into the current landscape and future trends. He also highlights the importance of compute and the transitory nature of AI coding, explored further in part two of his essay for Principal AI Coding members.
{# 
{#