Pi Reverses Its Anti-MCP Stance: 'The World Is Not Static'

The most vocal MCP holdout in the coding-agent world has surrendered, and the surrender note is worth reading. Pi, the coding agent from Earendil Engineering whose team (including Mario Zechner) was openly dismissive of Model Context Protocol, added MCP support to its core and published a September 29 post explaining the reversal under the title “You Said No MCP!” The one-line thesis: “the world is not static.” MCP itself improved, and the infrastructure Pi had to build for integration turned out to be useful beyond it. For a protocol that half the agent ecosystem adopted and the other half argued about, this is the clearest signal yet that the debate phase is over.

Why the Alternatives Failed

The interesting part of the post is the engineering honesty about what Pi tried first. An endorsed community extension (the pi-mcp-adapter) was insufficient, and so was the more radical option: skipping MCP entirely and using “Codemode,” Pi’s pattern of having a sandboxed JavaScript runtime orchestrate tool calls on the trusted harness side. The blocker was metadata. Deciding whether a given tool should be exposed to the LLM directly or routed only through the Codemode sandbox requires tool descriptions and capability metadata that Pi’s own loadout did not carry, and building that metadata layer for MCP-shaped tools required core changes anyway. Once the core work was unavoidable, embracing the protocol and shaping it beat forking around it, in the team’s words: “the best way to positively influence something is to embrace it.”

What Pi Wants MCP to Become

The post is also a requirements document for MCP’s next revision. Pi’s team argues MCP should work “much closer to OpenAPI with intelligent tool discovery”: tools returning structured, machine-readable data rather than token-optimized text dumps, with discovery smart enough that an agent loads only relevant tools instead of every attached server’s full catalog. They also name MCP’s biggest remaining flaw as composability (“it’s hard to compose”), while noting the problem has partly moved from the protocol to servers and harnesses. Their demonstration stitches the Linear MCP server to TypeSafe’s Jev classifier via Cloudflare Workers AI inside the Codemode sandbox, rating frustration across 167 open issues (156 neutral, 11 mildly frustrated, none highly) and storing the result under a session key for reuse, a small but complete picture of what compositional agent tooling looks like when the orchestration is code rather than prompt.

The Reversal Is the Standard Playing Out

Pi’s arc mirrors the industry’s in miniature: skepticism, an attempt to route around the standard, and eventual adoption once the integration tax exceeded the ideological cost. The lesson for tooling teams is about where standards come from: MCP won not because a committee mandated it but because the alternative, maintaining bespoke integrations per tool per agent, stopped being defensible as the tool ecosystem grew. The remaining open question Pi raises is real, though: protocol adoption settles the transport and discovery layers, while the hard problems (composability, token-efficient tool results, security boundaries of the kind the MCP-38 taxonomy catalogued for MCP systems) are still being invented.

What to Watch

First, whether MCP’s stewards absorb Pi’s OpenAPI-style discovery proposal, because intelligent tool loading is the difference between MCP scaling to hundreds of servers and drowning every agent’s context window. Second, whether Codemode-style JS orchestration becomes a standard harness pattern; if it does, the security boundary shifts decisively to the sandbox, and sandbox-escape research gets a new target. Third, the last holdouts: every agent that still ships a bespoke tool format is now making a choice that Pi’s post makes look increasingly expensive, and the next reversal announcements will tell us when MCP coverage is effectively total.