Pi.dev has reversed its previous stance on Model Context Protocol (MCP) support, integrating it directly into the platform’s core despite earlier public dismissals of the technology.
According to the Earendil blog, Pi previously featured “a proud declaration that Pi does not support MCP” on its website and team members made dismissive statements about MCP in podcasts and posts. However, users upgrading to the latest version of Pi now find MCP listed as a supported feature.
The shift reflects evolving perspectives on MCP’s maturation over the past year. Rather than implementing MCP as an extension—which Pi’s ecosystem already supports—the team decided to integrate it at the core level after reconsidering the platform’s broader architecture needs.
According to the post, the decision hinged partly on changes that would benefit both MCP and other functionality like Jev integration. “Ultimately what Pi needs is quite similar to what MCP needs: a sandbox to play with in the form of an interpreter,” the post explains.
The integration includes a new feature called Codemode, described as a mechanism for orchestrating and coordinating tool calls. Codemode operates in a trusted environment where the harness runs, rather than in untrusted sandboxes where most tools execute. It uses JavaScript to allow agents more flexibility in sequencing tool calls and combining results, with state maintained in session transcripts rather than the file system.
The team acknowledged persistent challenges with MCP. “The biggest issue with MCP continues to be that it’s hard to compose,” the post states, even with Codemode. The post attributes this partly to how many existing MCP servers are built—typically optimized around tools dumped directly into context for token efficiency.
Pi’s approach repositions MCP conceptually closer to OpenAPI with intelligent tool discovery, requiring tools to return structured data and be discoverable through documentation. The post notes that while “quite a few” improvements remain needed in the MCP ecosystem, the team chose to embrace and help shape the protocol rather than remain “on the sidelines.”
The integration demonstrates how Pi is adapting to newer language model capabilities, including deferred tool loading and mid-conversation system message changes. Codemode can be automatically loaded with MCP configuration or added separately, enabling applications like combining Linear MCP and Jev to analyze issue tracker data without consuming additional context.
Key facts
- Pi.dev previously declared it did not support MCP but now integrates it into core
- The change reflects MCP’s evolution and Pi’s need for better tool orchestration
- Codemode, a new sandbox feature, allows agents to sequence and combine tool calls using JavaScript
- MCP tools in Pi now follow an OpenAPI-style approach with structured data returns
- The integration supports newer LLM capabilities like deferred tool loading and mid-conversation system changes