MCP sucks honestly It eats too much context window and you have to toggle it on and off and the auth sucks I got sick of Claude in Chrome via MCP and vibe coded a CLI wrapper for Playwright tonight in 30 minutes only for my team to tell me Vercel already did it lmao But it worked 100x better and was like 100LOC as a CLI

🚄DSPy Code: 🎉 v0.1.1: Faster, Smarter, Friendlier The releases targeted for the uv users and faster response with managed RAG and Context. 🧰 First-class UV support, including project-root venv detection and uv-based install flows. 🧠 RAG & speed controls via /fast-mode, /disable-rag, /enable-rag, and /status so you can trade quality vs speed. Checkout latest DSPy Code: ⚡️uv pip install dspy-code OR add to existing project ⏯ uv add dspy-code #DSPyCode #uv #DSPy #GEPA #AgentOptimization

Recently, Anthropic anthropic.com/engineering/co… and Cloudflare blog.cloudflare.com/code-mod… released two blog posts that discuss a more efficient way for agents to interact with MCP servers, called Code Mode. There are three key issues when agents interact with MCP servers traditionally: - Context flooding - All tool definitions are loaded upfront, including ones that might not be necessary for a certain task. - Sequential execution overhead - Some operations require multiple tool calls in a chain. Normally, the agent must execute them sequentially and load intermediate return values into the context, wasting time and tokens (costing both time and money). - Code vs. tool calling - Models are better at writing code than calling tools directly. To solve these issues, they proposed a new method: instead of letting models perform direct tool calls to the MCP server, the client should allow the model to write code that calls the tools. This way, the model can write for loops and sequential operations using the tools, allowing for more efficient and faster execution. For example, if you ask an agent to rename all files in a folder to match a certain pattern, the traditional approach would require one tool call per file, wasting time and tokens. With Code Mode, the agent can write a simple for loop that calls the move_file tool from the filesystem MCP server, completing the entire task in one execution instead of dozens of sequential tool calls. We implemented Code Mode in mcp-use's (repo github.com/mcp-use/mcp-use ) MCPClient . All you need to do is define which servers you want your agent to use, enable code mode, and you're done! The client will expose two tools: - One that allows the agent to progressively discover which servers and tools are available - One that allows the agent to execute code in an environment where the MCP servers are available as Python modules (SDKs) Is this going against MCP? Not at all. MCP is the enabler of this approach. Code Mode can now be done over the network, with authentication, and with proper SDK documentation, all made possible by Model Context Protocol (MCP)'s standardized protocol. This approach can make your agent tens of times faster and more efficient. Hope you like it and have some improvements to propose :)

Connect any LLM to any MCP server! MCP-Use is the open source way to connect any LLM to any MCP server and build custom agents that have tool access, without using closed source or application clients. Build 100% local MCP clients.

The @mastra MCP Client class is one of the most feature rich plug-and-play implementations out there mastra.ai/en/reference/tools… What a great foundation that is. Choosing @mastra as the basis for PerfAgent just keeps paying off!