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 has made Model Context Protocol support part of Pi, despite previously saying the project would not support MCP. Earendil says the change pairs MCP tools with a JavaScript sandbox called Codemode, which can coordinate tool calls and also support integrations such as Jev.
Pi.dev has added Model Context Protocol (MCP) support to Pi, reversing an earlier position that the project would not support the protocol. In an explanation published by Earendil, the team says the change reflects improvements in MCP and the value of integrating it with Codemode, a JavaScript sandbox for coordinating tool calls.
Earendil says MCP had previously been available as an extension, but the team reconsidered whether it belonged in Pi’s core. Its explanation says the decision was not simply a judgment that MCP had improved: work needed to make MCP fit Pi also produced changes the team considers useful elsewhere, including making Jev easier to use within Pi.
In the new setup, MCP tools are exposed to a JavaScript sandbox. The team describes Codemode as a way for an agent to coordinate calls, choose their order and combine their results. Pi loads Codemode automatically when MCP is configured, according to the report; users can also add it as a default tool through configuration.
The report says Pi has recently worked on compatibility with models that support deferred tool loading, mid-conversation system messages and changes to reasoning level. It argues that Pi’s earlier tool configuration did not provide enough information to make MCP extensions work well with Codemode and these newer model capabilities. The team says it added options to configure tools as deferred or as Codemode-specific.
MCP Gets a Sandbox in Pi
The change gives Pi users a way to access MCP integrations while coordinating calls in a sandbox, rather than relying only on tools presented individually to the model. Earendil says that structure can let an agent combine tool calls and process their results in JavaScript, potentially reducing how much intermediate information enters the conversation. The report presents this as a design aim; it does not provide benchmark results or quantify any reduction in context use.
The decision also signals a change in the team’s relationship with MCP. Earendil says it wants to help shape how the protocol works in smaller agent harnesses. That matters to developers evaluating Pi because support in the core may make MCP configuration and tool access more direct, while Codemode creates a path for orchestration across multiple tools and services.
There are limits to what the announcement establishes. Earendil says MCP remains difficult to compose in practice, and attributes much of that difficulty to server design and differences between agent harnesses. The integration gives Pi a way to work with MCP; it does not establish that MCP servers have adopted the structured, discoverable approach the team favors.
From Rejection to Core Support
Earendil acknowledges that Pi’s earlier messaging was emphatic: visitors to pi.dev would have seen a declaration that Pi did not support MCP, and the team had made dismissive comments about the protocol in podcasts and a post by Mario. The source does not reproduce those earlier statements in full, but it describes the contrast with the current release as a reversal.
The team says it has been watching MCP over the past year and believes the protocol has changed. It also says that improvement alone would not have been sufficient reason to put MCP into Pi’s core. MCP had already been possible as an extension; the decision followed a wider reconsideration of Pi’s tool configuration and the role of a sandbox.
Earendil compares the intended direction for MCP to OpenAPI with intelligent tool discovery: tools should be findable through documentation and descriptions, and return structured data. It contrasts this with servers that return text and are designed for harnesses that place tools directly into the model’s context. The report says command-line tools work well partly because an agent can connect operations using shell commands, and argues that MCP can support similar coordination.
“The reason we brought MCP into the core is not just about how MCP has changed, but also because we found that the changes it would require were generally useful.”
— Earendil
Open Questions on MCP Use
The source does not specify when support shipped, which Pi versions include it, or whether every MCP server works with the new setup. It also gives no performance measurements, security evaluation or comparison of context use. Those details would help users judge the practical effect of the change.
Earendil says the protocol’s composition problems persist and points to server patterns and differences among harnesses as contributing factors. The report does not identify particular servers that have adopted structured results or improved discovery, nor does it say how broadly the team expects those practices to spread. Its example involving Linear and Jev illustrates a possible workflow, but the source does not provide independent evaluation of its results.
How Pi Shapes MCP Support
Pi users can configure MCP and use Codemode to coordinate tool calls; the report also says Codemode can be enabled as a default tool. Earendil says it wants to take part in the discussion about improving MCP for smaller harnesses, but does not announce a roadmap, named follow-up release or deadline.
The next evidence will come from implementation details and use in practice: which servers work well with Pi, how tool discovery and structured results are handled, and whether Codemode makes multi-tool workflows easier to compose. The source leaves those questions open.
Key Questions
What changed in Pi.dev’s MCP policy?
Pi has added MCP support to its core after previously saying the project did not support the protocol. Earendil says MCP had also been available as an extension.
What is Codemode?
Earendil describes Codemode as a JavaScript sandbox where an agent can coordinate tool calls, control their order and combine results. Pi loads it automatically when MCP is configured, according to the report.
Why did Pi add MCP to the core?
The team says MCP has changed and that the work needed to support it brought generally useful improvements to Pi’s tool configuration. It also says Codemode can help address some composition problems.
Does the change fix MCP’s composition problems?
Not completely, according to Earendil. The team says MCP remains hard to compose and points to server design and differences among agent harnesses as continuing issues.
When did MCP support become available?
The source describes the support as newly available but gives no release date or version number.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
