Pi 1.0 brings MCP into its core by batching tool calls in a JavaScript sandbox

According to Pi's team, many MCP servers were built for older harnesses that load every tool into the context window and then try to save tokens by returning plain text ^[15]. A dev.to writeup of the release describes how Codemode avoids that. The model writes one JavaScript script that orders the calls and merges the results, then returns only the final answer, instead of calling tools one at a time through a limited context window ^[10]. The writeup, by Nokka, says it was drafted by deepseek-v4.1-flash through Hermes Agent under human review ^[20].

A harness can execute a tool call in one of two places: where bash runs, or where the agent loop runs ^[8]. The loop usually sits in a trusted environment. Tools usually sit in a less trusted sandbox ^[8]. Codemode runs on the loop's side ^[9]. That puts the security argument on the choice of language. The team says a small version of JavaScript compiles to a WASM binary, and that this gives reasonable protection against risk ^[12]. Codemode keeps its state in the conversation log, not the filesystem ^[11]. We think that is good engineering. The writeup notes the design leaves no trace on the machine and lets a user look back at what the script did ^[11].

Pi 1.0 shipped on October 1, 2026 ^[2], and the site now reads "Now with MCP+Codemode" ^[3]. The change was visible. The launch post claims hundreds of thousands of weekly users without giving a number ^[23], and the 1.0 thread on Hacker News reached first place with 1,687 points and 609 comments ^[19]. The earlier skepticism was on the record too, including an older article by Mario Zechner asking what working without MCP would look like ^[4].

The team's other reasons concern Pi's own design. The restructuring MCP required also made Jev and classifier models easier to use in Pi ^[6]. And Pi's needs, the team wrote, turned out to resemble MCP's: a sandbox to play in, in the form of an interpreter ^[7].

In our view, the evidence supports a narrow conclusion. Pi took MCP into its core once it had an interpreter that keeps tool calls away from the model's context. The evidence covers one harness, and its team gives architectural reasons for the move. The team also stops short of calling the protocol fixed. It says MCP's biggest problem, that it is hard to compose, remains ^[14]. It wants MCP to move closer to OpenAPI and return structured data ^[16].

Size is the other test. Pi is a TypeScript project under the MIT license ^[1], pitched as small enough to understand whole and open to extension at almost every layer ^[21]. The team puts the Pi Durable code, excluding tests, at about 15,000 lines. That is roughly 150,000 tokens for GPT-family models and 250,000 for Claude, sized so an agent can read the harness it runs in ^[17]. Per line, that comes to 10 tokens against about 17, so the same self-inspection costs about two-thirds more on Claude ^[22]. The writeup does not say how many lines MCP and Codemode added to the core.