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 moved MCP support into the core of its AI coding agent after previously saying it would not support the protocol. The company says its implementation uses a JavaScript sandbox, called Codemode, to orchestrate tool calls and address some limits it sees in existing MCP integrations. It says MCP servers and usage patterns still need improvement.
Pi.dev has added Model Context Protocol (MCP) support to its AI agent, reversing an earlier position against the protocol. In a report published by Earendil, the team behind Pi said the change followed a reassessment of MCP and the addition of Codemode, a JavaScript sandbox for coordinating tool calls. The shift matters to Pi users who want to connect the agent to external tools and services through MCP.
Earendil said Pi users who upgrade will find MCP as a core feature. The company had previously said on Pi.dev and in podcast discussions that Pi would not support MCP; the report also points to an earlier post by Mario. An MCP extension for Pi existed before the functionality moved into the core, according to the report.
The team said its decision was not simply a response to changes in MCP. It found that the work needed to support the protocol could also make other capabilities, including access to Jev, easier to use within Pi. Earendil described Pi’s need as a way for an agent to compose tool calls inside an interpreter-like environment.
In Pi’s implementation, MCP tools are exposed to a JavaScript sandbox. The report says Codemode can coordinate calls and combine their results, with its state held in the session transcript. Pi loads Codemode automatically when MCP is configured, Earendil said, and users can also add it as a default tool.
MCP Moves Into Pi’s Core
The change gives Pi users a built-in route to MCP-connected services, while signaling that the team sees value in shaping the protocol’s use rather than leaving it to extensions. Earendil said it wants MCP to work well in small agent harnesses and to encourage better patterns among servers and tool integrations.
Codemode changes how the agent works with those tools. Instead of relying only on a harness to present tools individually, the agent can use JavaScript to arrange calls and combine structured results. That could help with multi-step tasks that draw on several services, though the report does not provide independent performance data or a comparison with other approaches.
Why Pi Reconsidered MCP
Pi’s earlier objection was tied to shortcomings the team saw in MCP and the servers built for it. Earendil said tool composition remains difficult, even with Codemode-like approaches, and that many servers are designed for systems that place large tool lists directly into the model’s context. Some servers return text to reduce token use, rather than structured data that can be reliably combined.
The company now describes its preferred direction as closer to OpenAPI paired with intelligent tool discovery: tools should return structured data and be discoverable through their descriptions. It also said Pi’s tool setup needed to account for newer models that support deferred tool loading, system messages during conversations and changes to reasoning level. The team said existing metadata was not sufficient to make MCP tools work cleanly with Codemode, prompting changes to how tools can be deferred or reserved for Codemode.
“The biggest issue with MCP continues to be that it’s hard to compose.”
— Earendil, in its report about Pi.dev
Open Questions on MCP Quality
The report does not specify when the core support shipped, which Pi versions include it, or what configuration requirements apply beyond enabling MCP and Codemode. It also gives no benchmarks, security evaluation or user data showing how the implementation performs on real tasks.
Earendil said MCP servers and the patterns used to build them still leave room for improvement. The report does not identify which servers return structured data, how broad Pi’s compatibility is, or whether the team’s preferred discovery and composition patterns will be adopted elsewhere.
How Pi Plans to Shape MCP
Earendil said it intends to take part in the discussion about improving MCP and making it work for smaller agent harnesses. For Pi users, the next practical step described in the report is to configure MCP; Codemode will then load automatically, or it can be set as a default tool.
The report does not name a release schedule or a specific follow-up milestone. Further details on supported servers, configuration, and how the team will evaluate the new setup remain to be announced.
Key Questions
What changed at Pi.dev?
Pi.dev added MCP support to Pi’s core after previously saying the agent would not support the protocol. An MCP extension had existed before the feature moved into the core, according to Earendil.
What is Codemode?
Codemode is a JavaScript sandbox that lets an agent coordinate tool calls and combine their results. Earendil says its state is held in the session transcript.
Why did Pi.dev change its position?
Earendil said its view of MCP had changed and that the work required for the integration could also support other useful features. The team also said MCP with Codemode could address some issues with composing tool calls.
Does Pi.dev say MCP’s problems are solved?
No. Earendil said composition remains difficult and that many MCP servers and usage patterns still need improvement. It described its implementation as one way to work with the protocol and influence its development.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
