Pi.dev Coding Agent Skips Anthropic's MCP, Igniting Debate Over Protocol's Limits

A blog post arguing that the coding agent Pi.dev should not adopt the Model Context Protocol has put a fresh question in front of agent builders: when does a shared standard stop earning its keep? The piece, published at earendil.com under the title "You Said No MCP," holds that the industry's push toward one universal tool-calling layer may be a poor fit for agents with narrow jobs. It reached the front page of Hacker News, where it collected 51 points and a 17-comment thread that split over the same tradeoff.

MCP arrived from Anthropic in late 2024 as a way for models to reach outside tools and data through one interface rather than a custom connector per API. Since then, OpenAI, Google and a broad set of developer-tool vendors have announced support, moving the protocol from a side project to something close to a default. That consolidation is the backdrop for the Pi.dev argument, which questions whether the same abstraction suits a coding agent that spends its time reading files, running commands and touching version control.

The objection laid out in the post is concrete rather than philosophical. A general-purpose protocol brings schemas, server processes and a transport layer along with it, and for a tightly scoped coding agent those pieces can add delay and new ways to fail between the model and the code it is editing. Terminals, file systems and Git already expose stable command-line interfaces a model can call directly, so stacking an abstraction on top is a choice that has to justify itself. Commenters who backed the post made essentially that point: Pi.dev's constraints differ enough from a chat assistant's that importing MCP wholesale could be the wrong call.

Not everyone in the thread agreed. Some respondents argued that standardization is exactly what allowed earlier tooling ecosystems to scale, and that a team opting out signs up to rebuild integrations that already exist elsewhere. That camp treats MCP as infrastructure rather than overhead, and sees divergence as a cost paid later. The disagreement is the story here, not the comment count, which is small even by Hacker News standards.

What gives the post weight is its timing. A year ago the debate would have been about whether MCP would catch on at all; now that it has, the sharper engineering question is where it stops being the right fit. Pi.dev is effectively being offered as a test case for that boundary, and the thread reads as an early referendum on it.

The next signal will be whether the no-MCP stance survives contact with a shipped product. If Pi.dev's integrations hold up without the protocol, teams facing the same decision get a real data point; if the project later adds MCP support, the argument becomes a record of a position that did not last. Either outcome tells agent developers more than the thread itself can.