TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Pi.dev now includes Model Context Protocol (MCP) support, reversing its earlier position that the agent would not support it. The developers say MCP tools run through Pi’s Codemode JavaScript sandbox, which can coordinate calls and combine their results. They say the change also supports broader improvements to Pi’s tool system; details such as release timing and configuration requirements are not specified in the source report.
Pi.dev now supports the Model Context Protocol (MCP) as part of the core product, reversing its earlier public position against adding the protocol. In a report published by Earendil, the Pi developers say the integration routes MCP tools through Codemode, a JavaScript sandbox that lets an agent coordinate tool calls and work with their results.
The developers say MCP support was previously available as an extension, but is now part of Pi itself. They describe the decision as a response to changes in MCP and to broader work on how Pi handles tools. According to the report, those changes also make it easier to use Jev within Pi. The source does not specify when the core feature was released or which Pi version includes it.
In Pi’s design, Codemode gives the agent a place to sequence tool calls and combine their outputs using JavaScript. The developers say MCP is exposed to that sandbox, rather than relying only on a conventional setup that presents every tool directly in the model’s context. They describe the sandbox as part of the harness, with its state kept in the session transcript.
The report gives an example involving Linear and Jev: Pi can retrieve open issues, send thread content for tone classification, and combine the results. That example illustrates the intended workflow, but it is a demonstration from the developers, not an independent performance evaluation. The source does not provide benchmark results or a security assessment of the implementation.
Pi Brings Tool Calls Into Codemode
The change matters to Pi users who need agents to work across several services. Coordinating tool calls in a sandbox can let an agent retrieve information, pass it to another tool, and combine the outputs without placing every tool result into the conversation as plain text. The developers say this approach can reduce unnecessary context use, though the report gives no measurements of token savings.
Making MCP a core feature also gives Pi’s developers a direct role in shaping how MCP tools work within a smaller agent harness. Their stated aim is to encourage tools with structured results and useful descriptions, which could make them easier to discover and combine. Whether MCP servers adopt those patterns, and how much the integration improves users’ workflows, remain open questions.
Why Pi Changed Its MCP Position
Pi had publicly said it did not support MCP, including on its website and in discussions about the product, according to the Earendil report. The developers acknowledge that the protocol could have remained an extension; they say the decision to put it in the core followed a reassessment of both MCP and changes that would benefit Pi more broadly.
The report identifies tool configuration as one such change. Newer models can support deferred tool loading, while Pi’s existing tool setup did not yet make it easy to decide whether a tool should be available directly to the model or only through Codemode. The developers say improving that configuration was needed for the MCP integration and could help Pi make better use of newer model capabilities.
“The biggest issue with MCP continues to be that it’s hard to compose.”
— Pi developers, in the Earendil report
Release Details Remain Unspecified
The source report does not give the release date, the Pi version required, or a detailed setup guide for MCP servers. It says Codemode loads automatically when MCP is configured, and that users can also add Codemode as a default tool, but it does not spell out the steps or identify any compatibility limits.
It is also unclear how the implementation handles permissions and failures across multiple tools, or whether all MCP servers can return the structured data the developers favor. The report offers a developer-provided example but no independent tests, usage figures, security review, or comparison of context use. Claims about improved composition and reduced context overhead should be read as the developers’ rationale, not measured outcomes.
Pi’s Tool Configuration Evolves
Pi’s next steps depend on how its tool configuration and MCP support develop. The developers say Pi needed options to mark tools as deferred or Codemode-specific, and present that work as part of adapting the product to models with newer tool-loading capabilities. The report does not announce a dated roadmap or a specific upcoming release.
For users, the immediate practical step described is to configure MCP in Pi; Codemode then loads automatically, or can be added as a default tool. Further details about setup, supported servers, and future changes would clarify how the new core feature works in everyday use.
Key Questions
What changed at Pi.dev?
Pi now supports MCP in its core product. The developers say MCP had previously been available as an extension.
What is Codemode?
Codemode is a JavaScript sandbox in Pi’s harness that lets the agent coordinate tool calls, choose their order, and combine their results.
Why did Pi add MCP after opposing it?
The developers say MCP has changed and that the work needed to support it also improves Pi’s tool configuration. They also want to help shape how MCP works in smaller agent harnesses.
Does the report show that MCP uses fewer tokens?
No. The developers say the approach can avoid wasting context, but the report includes no token-use measurements or benchmarks.
When was the feature released?
The source report does not give a release date or identify the Pi version that introduced core MCP support.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
