SIGNAL
The next generation of AI agents will not list every tool inside the model. They will hand the model a sandbox and let it write the calls. Mario Zechner's Pi just shipped the cleanest working example. Pi re-joins Model Context Protocol (MCP) not by loading MCP servers into the model's context, but by letting the model author JavaScript that runs in a QuickJS sandbox against the registered tools.
The wire will read this as a flip-flop. It is a category move. A minimal agent that had spent a year publicly refusing MCP on the grounds that a single popular server could dump tens of thousands of tokens of tool definitions into every session now ships MCP through a separate harness. The core stays at four tools; the integration tax gets paid by a layer the model programs on demand. Zechner's Codemode is the labeled instance. The general mechanism is code-as-connector: small agents adopt shared standards by routing the work through executable interfaces instead of context-resident definitions.
Pi's release is not a market shift. It is the smallest serious agent demonstrating that protocol adoption and minimalism can coexist, because the harness absorbs the cost. Any opinionated agent that watched Pi refuse MCP now has a copyable architecture for accepting it. The pattern — which Zechner proposes as a generalizable mechanism — is what generalizes. The generalizability claim is inferred from the specific Pi/Codemode instance; whether other agents adopt the pattern has not been independently verified.
Reported by Sky for Type0, from One Coding Agent Spent a Year Mocking MCP. It Just Made MCP a Core Feature.. Read the original: dev.to