AI coding tool Pi has formally added built-in support for MCP in version 0.99.0, while also changing the wording on its website from "No MCP" to "Now with MCP+Codemode." The move stands out because Pi had previously criticized the standard way MCP is commonly connected to models. Mario Zechner, the author cited in the report, said tests he ran last year showed that the descriptions for 21 tools in Playwright MCP used about 13,700 tokens, while 26 tools in Chrome DevTools MCP used about 18,000 tokens. In his view, exposing too many tools wastes context, makes composition harder, and pushes large volumes of returned content directly into the model context. Pi's new setup keeps the general MCP interface but changes how tools are surfaced. Instead of exposing all MCP tools to the model by default, Pi hands them to Codemode. The model can then write JavaScript to call tools in parallel, combine and filter them, and send back only the results that are needed. Tools can also be loaded later instead of being declared up front. Pi team member Earendil has not withdrawn the earlier criticism, saying many MCP servers still pack in large numbers of tools and text outputs, and that the tool-combination issue has not been fully solved.
AI coding tool Pi has added built-in MCP support in version 0.99.0. Its website also dropped the old "No MCP" wording and replaced it with "Now with MCP+Codemode."
That shift is notable because Pi had openly criticized the common way MCP is usually connected. According to the report, author Mario Zechner said tests he ran last year found that the descriptions for 21 tools in Playwright MCP took about 13,700 tokens, while 26 tools in Chrome DevTools MCP took about 18,000 tokens.
He argued that too many tools waste context and are hard to compose. Large amounts of returned content also go straight into the model context.
Pi is now using a different approach. MCP tools do not need to be fully exposed to the model by default and are instead handled by Codemode. The model can write JavaScript to call tools in parallel, combine them, and filter outputs, then return only the results it needs to the context. Tools can also be loaded on demand rather than being declared at the start.
Pi team member Earendil has not walked back the earlier criticism. The team's view is that many MCP servers still stuff in large numbers of tools and text results, and that the problem of tool composition has not been completely solved.
Pi now keeps MCP's general interface while using Codemode to avoid the usage pattern it had been most dissatisfied with before.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
