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, a coding agent that publicly rejected MCP support, has reversed course and shipped MCP as core functionality. A post by Earendil explains the reversal: modern MCP changed, and the underlying work also enables a JavaScript orchestration sandbox called Codemode.
Pi.dev has reversed its public stance against the Model Context Protocol (MCP) and shipped MCP as a supported core feature, according to a post by Earendil titled “You Said No MCP.” The reversal is notable because Pi’s website previously carried what the author calls “a proud declaration that Pi does not support MCP,” and the team had made dismissive statements about the protocol on podcasts and in posts. Users who upgrade to Pi will now find MCP available alongside a new JavaScript orchestration sandbox called Codemode.
The author frames the decision around two factors. First, “the MCP of today is not the MCP of yesteryear” — the protocol has improved over the past year while the team watched it. Second, and described as the stronger reason, the engineering changes needed to support MCP well turned out to be generally useful beyond MCP itself. According to the post, those changes “also enable the use of Jev more easily within Pi,” referencing a classifier model used for analysis tasks.
MCP had previously existed as an extension rather than a core capability. Earendil confirms this was a deliberate rethinking, not a forced migration: “That MCP is now part of the core is a result of us putting our heads together and rethinking it.” The team’s stated philosophy is that “what Pi needs is quite similar to what MCP needs: a sandbox to play with in the form of an interpreter.”
The implementation pairs MCP with Codemode, a sandbox that runs on the harness side of the agent loop rather than in the less-trusted tool-execution environment. Codemode lets an agent issue and combine tool calls using JavaScript, with its state maintained in the session transcript rather than the file system. In Pi, Codemode is loaded automatically when MCP is configured. The post demonstrates combining the Linear MCP server with Jev to identify “the 20 most frustrated commenters” on an issue tracker, which the author says can be done “without wasting any context at all.”
Why the Reversal Matters for MCP’s Trajectory
Pi was one of the most visible holdouts against MCP, so its adoption signals how far the protocol’s position has hardened across the agent-tooling ecosystem. But the post is not an unqualified endorsement: Earendil writes that “the biggest issue with MCP continues to be that it’s hard to compose,” and argues that many MCP servers are still built for harnesses that dump tools into context and return text optimized for token efficiency rather than structured data.
The team’s proposed direction is that MCP “should be much closer to OpenAPI with intelligent tool discovery” — tools returning structured data and being discoverable through documentation. Earendil states the motivation plainly: “We believe the best way to positively influence something is to embrace it,” positioning Pi as a participant shaping how MCP works in small harnesses rather than a critic on the sidelines.
For Pi users, the practical payoff is the combination of MCP servers with Codemode’s composition abilities, plus better support for newer model capabilities such as deferred tool loading, mid-conversation system messages, and reasoning level changes, which the team says its tool loadout had not yet scaled to.
Pi’s History of Rejecting MCP
: “Before this change, Pi’s website explicitly stated the project did not support MCP, and team members — including one identified as Mario — had published dismissive commentary about the protocol and repeated that position in podcast appearances, according to the post. MCP was nonetheless available in some form as an extension, which Earendil acknowledges could have remained the arrangement.
The technical background matters for understanding the reversal. Pi had done recent work to accommodate modern LLM features, but its tool loadout lacked the metadata to let an extension decide whether a tool should be exposed to the full model or only to a Codemode-style sandbox. Wiring up that metadata is what made core MCP support viable, the post explains, and it also addresses composition problems the team says MCP has traditionally had.
“The MCP of today is not the MCP of yesteryear.”
— Earendil, in the post ‘You Said No MCP’
Open Questions About the New Setup
Several points remain unclear from the post alone. It does not specify which MCP servers have been tested with Pi’s implementation, or how the structured-data vision will work with existing servers that return plain text. The full Codemode API and the truncated Jev example are not documented in the post, so the exact capabilities available to users today are not fully detailed. The post also does not give a release version or timeline beyond saying MCP is available to users who upgrade, and it does not address whether the old MCP extension will be deprecated. Finally, whether the broader MCP ecosystem will move toward the OpenAPI-style, discoverability-focused model the team advocates is an open question the post itself leaves to future conversation.
Pi’s Role in Shaping MCP Servers
According to the post, the team intends to participate in shaping how MCP servers and patterns evolve so they work well in small harnesses like Pi. Users can enable Codemode by asking Pi to reconfigure itself, with or without MCP configured. Documentation of the Codemode API, broader server compatibility, and follow-up on the structured-data and tool-discovery proposals are the natural next milestones, though no dates were given in the post.
Key Questions
Did Pi.dev previously support MCP?
Only as an extension. Pi’s website and team commentary publicly rejected MCP as a core feature, though an extension existed. MCP is now part of the core after a rethink described in the Earendil post.
What is Codemode in Pi?
Codemode is a JavaScript sandbox that runs on the harness side of the agent loop. It lets an agent orchestrate and combine tool calls with JavaScript, with state kept in the session transcript rather than the file system. It loads automatically when MCP is configured.
Why did Pi change its mind about MCP?
Per the post, two reasons: MCP itself improved over the past year, and the engineering changes needed for good MCP support also enabled other features like easier use of Jev and better scaling to modern model capabilities such as deferred tool loading.
Does Pi’s team still have criticisms of MCP?
Yes. The post says MCP remains hard to compose and that many servers still return text optimized for token efficiency instead of structured data. The team advocates an approach closer to OpenAPI with intelligent tool discovery.
How do I enable MCP and Codemode in Pi?
According to the post, upgrading to the current Pi release includes core MCP support, and Codemode is automatically loaded when MCP is configured. Users can also ask Pi to reconfigure itself to enable Codemode as a default tool.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
