Thread by meng shao (@shao__meng)

[Open Source Tutorial] Open Source Model + Mature Agent Framework + Tools => Replicating Claude Code-Level AI Agent Key technical components and open-source model: DeepSeek-V3.2 • Mature Agent Framework: Claude Agents SDK Tools and data: MongoDB MCP Server Overall architecture: Model → Agents SDK → MongoDB tool → Database, realizing closed-loop interaction. Core project concept: A powerful combination of three key technologies. Building an intelligent agent that can understand human speech and automatically manipulate databases integrates three technologies: 1. The Brain – DeepSeek v3.2: The "brain swap" operation involves modifying the API Base URL to make the Claude Agents SDK mistakenly believe it is calling the Claude model, when in fact it is calling DeepSeek v3.2. This has become a standard operation for LLM APIs after the OpenAI API. 2. Skeleton – Claude Agents SDK: Reason for selection: LangChain or OpenAI SDK were not selected because the Claude Agents SDK provides mature "scaffolding" (such as sub-agent management, MCP support, etc.) needed to build complex agents, which are the core technologies driving Claude Code. 3. Hand-eye coordination – MongoDB MCP Server: Technical highlights: Employing the MCP protocol, AI can perform queries, analyze schemas, and even write data in a standardized manner through MongoDB's MCP server, without the need for complex glue code. Architectural essence: Combating "brain fog" with "sub-agents" The most technically challenging part of the tutorial. The author raises a key issue: Context Rot. Even if a model claims to support 200k+ contexts, it becomes unintelligent and confuses tools when given too much input. Solution: Divide and Conquer (Subagents) Instead of using a single all-powerful agent, the tutorial constructs three specialized sub-agents, each responsible for only a portion of the MongoDB MCP toolset: • Reader Agent: Only responsible for reading (retrieving data). • Writer Agent: Only responsible for writing (adding, deleting, and modifying). • Query Agent: Responsible for finding relevant data based on fuzzy commands. Advantages: By limiting the field of vision and toolbox of each agent, DeepSeek greatly reduces the probability of making mistakes, ensuring the accuracy of operations. Practical Value: From "Toy" to "Tool" The tutorial not only demonstrates a simple demo like "how many movies are there", but also provides a highly relevant real-world example: • Data migration and analysis: The script demonstrates how to import real-world data (model statistics, dataset popularity, etc.) from Hugging Face Hub into MongoDB. • Complex queries: After importing, you can directly ask the Agent: "What are the 10 most popular models on Hugging Face?" The Agent will automatically generate an aggregate query and extract the answer from the database. Summary • Model Demystification: You don't need to rely on expensive closed-source models (such as Claude Opus 4.5). DeepSeek v3.2, with its good architecture, is perfectly capable of handling complex tasks. • MCP adoption: Connecting to databases via the MCP protocol will become the standard, significantly lowering the barrier to entry for developing AI applications. • Architecture First: Compared to pursuing longer contexts, the "main agent + specialized sub-agents" architecture is a stable solution for solving complex problems. Original tutorial