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 now includes support for the Model Context Protocol (MCP), reversing its earlier public stance against the standard. Earendil says the decision followed changes in MCP and Pi’s own tool system, with MCP tools now exposed through a JavaScript sandbox called Codemode.
Pi.dev now supports MCP in its core product, reversing earlier public statements that the AI agent harness would not support the Model Context Protocol. In a report titled “You Said No MCP,” Earendil says the change reflects both developments in MCP and updates to Pi’s tool system that make sandboxed tool orchestration useful beyond MCP itself.
Earendil says MCP had previously been available as an extension, but the team reconsidered whether it belonged in Pi’s core. The report says the decision was not simply a response to changes in MCP: work to support it also produced changes that make other capabilities, including Jev, easier to use within Pi.
In Pi, MCP tools are exposed to a JavaScript sandbox called Codemode. The report describes Codemode as a way for an agent to coordinate tool calls, choose their order and combine results in JavaScript. Its state is kept in the session transcript rather than the file system. Earendil says Codemode loads automatically when MCP is configured, and can also be added as a default tool.
The report says MCP remains difficult to compose across tools. Earendil attributes much of that difficulty to the servers and harness patterns in use: many servers, it says, were designed for systems that place tools directly into model context and return text. The team says it prefers a model closer to OpenAPI, with structured tool results and discovery through documentation and descriptions.
Pi Makes Tool Orchestration Central
Moving MCP into Pi’s core gives users a built-in route to connect external tools, while Codemode is intended to let the agent combine those tools without placing every individual call directly into its working context. Earendil says this can reduce context use and allow more flexible sequences of calls.
The change also marks a shift in how Pi’s developers intend to engage with MCP. The report says they want to help shape the protocol and server patterns for smaller agent harnesses. That is the team’s stated aim; the source does not provide independent measurements of context savings or evidence that the integration resolves MCP’s broader composition problems.
From Extension to Core Support
Pi’s earlier position, as described by Earendil, was that it would not support MCP, and MCP had been handled through an extension. The team says it had watched the protocol over the preceding year and found that it had changed, but says that alone did not justify adding it to the core.
Another factor was Pi’s recent work to support models with features such as deferred tool loading, mid-conversation system messages and adjustable reasoning levels. Earendil says Pi’s tool loadout had not yet been adapted to those newer capabilities. Codemode required clearer configuration of which tools are available to the model and which are available only within the sandbox, prompting broader tool-system changes.
The report compares Codemode with running commands in a shell or through a harness’s agent loop. It says Codemode runs on the harness side and is meant for orchestrating calls, while the tools themselves may run in less trusted environments. Earendil cites small JavaScript runtimes that can ship as WebAssembly binaries as one reason for choosing JavaScript.
“The biggest issue with MCP continues to be that it’s hard to compose.”
— Earendil, in “You Said No MCP”
Open Questions on MCP Composition
The report does not specify when core MCP support was released, which Pi versions include it, or whether users need to change existing configurations. It also does not provide implementation details about Codemode’s security boundaries or independent evaluations of its protections. Earendil says tool composition remains a problem and points to server design and differences between harnesses, but does not identify particular servers or quantify how often users encounter these issues.
Further Work on Tools and Servers
Earendil says it wants to take part in improving MCP for smaller agent harnesses and argues for tools that return structured data and can be discovered through documentation. The report does not announce a delivery schedule or name a specific next release. Pi users can configure Codemode alongside MCP, according to the report, but further details about future changes remain unstated.
Key Questions
What changed at Pi.dev?
Pi.dev added MCP support to its core product after previously rejecting core support. Earendil says MCP had earlier been available through an extension.
What is Codemode?
Codemode is a JavaScript sandbox for coordinating and combining tool calls. Earendil says it runs on the harness side and keeps its state in the session transcript.
Why did Pi add MCP to the core?
Earendil cites changes in MCP and updates to Pi’s tool infrastructure. The report says those infrastructure changes also make it easier to use capabilities such as Jev.
Does the report say MCP’s composition problems are solved?
No. Earendil says MCP remains hard to compose and links the problem partly to server design and the way different harnesses use tools.
Source: hn
Evergreen bestsellers Picks
bestsellers
As an affiliate, we earn on qualifying purchases.
