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 supports the Model Context Protocol (MCP), reversing its earlier public position against the standard. The developers say MCP tools run through a JavaScript sandbox called Codemode, which lets agents coordinate tool calls; they also cite broader improvements to Pi’s tool-loading system.
Pi.dev has added support for the Model Context Protocol (MCP), reversing its earlier public stance against the standard. In an Earendil post titled “You Said No MCP,” the developers say the integration uses a JavaScript sandbox called Codemode to coordinate tool calls, a change that also supports broader improvements to how Pi handles tools.
The developers acknowledge that Pi’s website had stated that it did not support MCP and that they had made dismissive comments about the protocol in podcasts and a post by Mario. They say an MCP extension already existed for Pi, but that they later reconsidered whether the functionality should be built into the core. Their explanation is that both MCP and Pi benefit from changes to how tools are loaded and run.
In Pi, MCP tools are exposed to a JavaScript sandbox rather than being presented only as a list of calls for the model. The developers describe this setup as Codemode: a way for an agent to combine and sequence tool calls in JavaScript. Pi can load Codemode automatically when MCP is configured, or it can be added as a default tool, according to the post.
The post also points to Pi’s work on handling newer models that support deferred tool loading, mid-conversation system messages and changes to reasoning levels. The developers say Pi’s existing tool configuration did not yet make those capabilities work well with Codemode. They added options to mark tools as deferred or Codemode-specific, and say those options also help MCP integrations fit the system.
Why Pi Changed Its MCP Policy
The change brings a widely used integration standard into Pi’s core and gives the project a way to influence how MCP tools work in smaller agent harnesses. The developers say they want tools to be discoverable through documentation and descriptions and to return structured data. Those design goals could make it easier for an agent to combine tools without loading every tool description into the conversation at once.
Codemode also changes where coordination happens. The developers describe it as running alongside the harness, with its state kept in the session transcript rather than the file system. They say JavaScript is a practical choice because small JavaScript runtimes can be shipped as WebAssembly binaries and provide a degree of protection. The post does not provide security testing or comparative performance results for Pi’s implementation.
The authors say the same work can help with uses beyond MCP. As an example, they describe asking Pi to use Linear issue data with Jev, a classifier available through a provider, to identify frustrated commenters. This is an example offered by the developers; the post does not publish an evaluation of the result or quantify any reduction in context use.
From Extension to Core
Pi’s earlier position, as described by the Earendil authors, was that MCP did not belong in the product. The post says an MCP extension was available before the core integration, showing that Pi users could already connect the protocol through its extension system. The developers now argue that incorporating it into Pi is useful because the underlying tool-loading changes have applications elsewhere.
The authors do not say that MCP’s problems have disappeared. They identify tool composition as a persistent weakness and say many servers are designed for harnesses that insert tools directly into a model’s context. Some servers return text to reduce token use, they write, while Pi’s preferred direction is structured results and tool discovery. They also attribute some remaining limitations to the servers and to differences among harnesses, rather than to the protocol alone.
The post compares the desired experience to OpenAPI with intelligent tool discovery and notes that command-line tools can be combined efficiently through shell commands. The authors argue that MCP could support a similar pattern when tools are coordinated in a sandbox. That is their design rationale, not an independent assessment of MCP or a claim that every server currently supports it.
“The biggest issue with MCP continues to be that it’s hard to compose.”
— Earendil developers
Limits the Post Leaves Open
The post does not specify when the core support shipped, which Pi versions include it, or whether every MCP server works with the integration. It also gives no independent test results comparing Codemode with direct tool calls or with other harnesses. The authors’ example combining Linear and Jev illustrates a possible workflow, but the available source does not show a complete evaluation or measure its accuracy.
Security details are also limited. The developers describe Codemode as a sandbox and say compact JavaScript runtimes can provide reasonable protection, but do not outline threat assumptions, isolation guarantees or known limitations. It remains unclear how Pi handles the range of server behaviors the authors say can make MCP hard to compose.
How Pi Plans to Shape MCP
The developers say they intend to take part in the discussion about improving MCP servers and usage patterns, with a focus on smaller harnesses. The post does not name a release schedule, planned compatibility milestones or a specific next update. For users, the immediate next step is to configure MCP in Pi and, if needed, enable Codemode; the authors say Pi can load it automatically when MCP is configured.
Further details about supported servers, tool metadata and security behavior would help show how the integration works across different setups. Until those details or independent evaluations are available, the post establishes Pi’s change in direction and its design rationale, while leaving the breadth and practical results of the support open.
Key Questions
What changed at Pi.dev?
Pi.dev has added MCP support to its core after previously saying it did not support the protocol. An MCP extension existed before the core integration, according to the developers.
What is Codemode?
Codemode is a JavaScript sandbox that lets an agent coordinate and combine tool calls. Pi can load it when MCP is configured, or users can add it as a default tool, the post says.
Why did Pi’s developers change their position?
The developers say MCP has changed and that the work needed to integrate it also improves Pi’s tool-loading system. They cite support for deferred tools and Codemode-specific tools as part of that work.
Does Pi say MCP’s problems are solved?
No. The authors say tool composition remains difficult and point to limitations in MCP servers and in the ways different harnesses use them.
What is not yet clear about the release?
The source does not give a release date or version, a compatibility list, independent performance results or detailed security guarantees for Codemode.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
