Show HN: Dynamic Code Execution with MCP: A More Efficient Approach (github.com)

Anthropic’s Code Execution with MCP pattern — demonstrated in the codex-mcp example — shows how agents can call MCP tools dynamically at runtime instead of generating TypeScript files for every tool. Using Vercel AI SDK’s MCP runtime, agents can discover tools with list_mcp_tools(), fetch schemas on demand with get_mcp_tool_details('tool_name'), and execute in-memory snippets that call tools via an injected callMCPTool function. Snippets (created, listed, edited, executed, and stored in chat session data) run immediately with no filesystem imports, compilation, or distribution of thousands of generated files, eliminating type-generation, build-pipeline, synchronization, version-conflict, and disk‑space overheads while preserving progressive disclosure, privacy, and persistent session state. The tradeoff is protocol-level: MCP doesn’t guarantee tool output schemas, so chaining tool calls requires defensive coding — assumptions about structure, parsing/validation, and sometimes extra verification calls. That limitation affects any MCP implementation, not just the dynamic approach. For practitioners, the takeaway is practical: dynamic, in-memory MCP execution dramatically reduces development and maintenance costs and keeps tools live-sync’d, but teams must add runtime validation and error handling to safely compose multi-tool workflows.