TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Pi.dev, which publicly refused to support the Model Context Protocol for years, has shipped MCP as core functionality. A post by Earendil explains the reversal: MCP improved, and integrating it required changes that benefit Pi’s broader architecture, including its new Codemode tool-orchestration sandbox.
Pi.dev has shipped the Model Context Protocol (MCP) as core functionality, reversing years of public refusal to support the protocol — including a declaration on its own website that it did not support MCP. In a post titled “You Said No MCP,” published by Earendil, the developers explain that the protocol has improved enough — and Pi’s needs converged with it enough — that keeping MCP out no longer made sense.
The authors acknowledge the about-face directly. Visitors to pi.dev previously found “a proud declaration that Pi does not support MCP,” and past podcast appearances and a post by Mario contained dismissive statements about the protocol. “And yet, if you upgrade to Pi you will find MCP is now a supported piece of functionality,” the post reads.
According to Earendil, two things drove the change. First, “the MCP of today is not the MCP of yesteryear” — the protocol has improved over the past year, which the team says it has been tracking. Second, and more decisive, the engineering work required to integrate MCP well turned out to be generally useful to Pi itself. The same changes, the post states, also make it easier to use Jev, a classifier model, within Pi.
MCP had previously existed as an extension, and the post concedes it could have stayed one. The decision to move it into the core came after the team “put our heads together and rethinking it.” A central piece of the new implementation is Codemode, a JavaScript sandbox that runs on the harness side — the more trusted side — of the agent, letting the model orchestrate and combine tool calls programmatically rather than dumping each tool result into the conversation context. In Pi, Codemode loads automatically when MCP is configured. The post demonstrates the combination with an example task — finding the most frustrated commenters on an issue tracker — where Pi pulls issues from the Linear MCP server and has Jev rate the tone of each thread, “without wasting any context at all.”
Why Pi Reversed Its MCP Position
The reversal matters for two reasons. For Pi users, it removes what had become a genuine limitation: MCP has become the de facto standard for connecting agents to external tools and services, and Pi’s refusal cut it off from that ecosystem. Support is now first-class, and the Codemode integration offers a different consumption model than most harnesses use.
More broadly, the post is a pointed critique of how MCP is currently used in practice. The authors argue the protocol’s biggest remaining problem is that “it’s hard to compose” — many MCP servers are still built for harnesses that dump tool descriptions into context and return plain text optimized for token efficiency. Earendil’s position is that MCP should look “much closer to OpenAPI with intelligent tool discovery”, with tools returning structured data and being discoverable through documentation. The team says it wants to shape that evolution: “We believe the best way to positively influence something is to embrace it.”
From Public Refusal to Core Integration
: “Pi’s anti-MCP stance was not quiet. Beyond the website declaration, team members — including Mario, who wrote a separate post about it — voiced skepticism on podcasts. During that period, MCP support existed only as an extension rather than built-in functionality.
Underneath the reversal, the post describes recent architectural work: Pi has been upgraded for newer model capabilities including deferred tool loading, mid-conversation system messages, and reasoning-level changes. But its tool loadout had not yet scaled to those capabilities. A plain MCP extension lacked the metadata to decide whether a tool should be exposed to the main model or only to Codemode — a gap the core integration now fills. Codemode itself, the post notes, could in principle use any language, but JavaScript was chosen because small versions can ship as WASM binaries with reasonable isolation guarantees.
“And yet, if you upgrade to Pi you will find MCP is now a supported piece of functionality. What happened? Things Change.”
— Earendil, blog post
Open Questions Around the New MCP Design
It is not clear when MCP support first shipped to users or which Pi version introduced it; the post does not give release dates. The authors themselves flag that the MCP ecosystem remains immature — many servers, they say, are still built for older harness patterns and do not return structured data, meaning the composed-tool-calling experience may vary widely depending on which servers a user connects. The post also does not detail how the deferred tool loading and Codemode-specific tool configuration will behave with specific models or providers, beyond the general design description.
Pi’s Role in Shaping MCP Servers
The team says it wants to “help shape” MCP to work well in small harnesses rather than watching from the sidelines, which suggests future work on server patterns and possibly tooling that encourages structured, discoverable MCP servers. Users can enable the functionality today: Codemode loads automatically when MCP is configured, or can be added as a default tool by asking Pi to reconfigure itself. The post also hints at broader uses for Codemode beyond MCP, such as combining it with provider-hosted models like Jev for analysis tasks without consuming context.
Key Questions
Did Pi.dev previously refuse to support MCP?
Yes. The pi.dev website carried a declaration that Pi does not support MCP, and team members made dismissive statements about the protocol on podcasts and in posts. MCP was available only as an extension before being moved into the core.
Why did Pi change its position on MCP?
According to Earendil, MCP itself improved over the past year, and the changes needed to integrate it well proved generally useful to Pi’s architecture — including making Jev easier to use and enabling better tool configuration for newer model capabilities.
What is Codemode in Pi?
Codemode is a JavaScript sandbox that runs on the harness side of the agent, where the trust level is higher. It lets the model orchestrate and combine tool calls programmatically, with state kept in the session transcript rather than the file system.
Does the Pi team still have criticisms of MCP?
Yes. The post states that MCP remains hard to compose, and blames many MCP servers for being built for harnesses that dump tools into context and return text. The team argues servers should return structured data and become discoverable like an OpenAPI-style interface.
How do users enable MCP in Pi?
MCP is now part of Pi’s core. Codemode loads automatically when MCP is configured, and it can also be added as a default tool by asking Pi to reconfigure itself.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
