MCP and the Future of AI (www.contraption.co)

Model Context Protocol (MCP) is emerging as a practical bridge between LLMs and the real apps people use: MCP servers expose callable “tools” (functions) that let agents search, fetch, paginate, and take actions across services. The author built an MCP for their blog (search, fetch, list_posts) and recounts how Codex connected to Honeycomb, Package Search, and GitHub MCPs to find an error trace, write a failing test, patch code, commit, and open a PR with minimal human intervention. Major platforms (Claude, ChatGPT, Cursor, VS Code) and many tools (Figma, GitHub, Notion, Stripe, AWS) now offer MCPs or community stand‑ins, enabling agents to access private, live data and perform compound workflows across systems. That capability matters because it lets agents run full OODA loops—observe logs/docs, orient across sources, decide with tests/heuristics, and act by updating dashboards or shipping code—unlocking real autonomy for developer and non‑developer tasks. But MCP today is developer‑centric and fragmented: manual config files, inconsistent UX, spotty authentication, and platform lock‑downs (e.g., ChatGPT’s restrictive setup). There are also security risks: MCP responses are injected into an LLM’s context and can poison chains of thought or carry malicious instructions. Expect MCP to remain the low‑level plumbing while model providers layer curated, app‑store‑like ecosystems for safer, one‑click adoption.