Earendil rebuilt Pi's agent core before adding the MCP support it once refused

On the wire, an MCP tool is a function announcement ^[6]. The server publishes a name, a description and a parameter schema in JSON Schema, and the model decides when to call it ^[6]. Client and server exchange JSON-RPC over a common transport ^[4]. The client lives inside the assistant, discovers the server's tools, resources and prompts when it connects, and chooses which ones the model sees on each turn ^[5]. The difference from provider function calling is that the announcement no longer depends on the model vendor or the host application ^[6].

Earendil's objection sat in the loop around that call, and a dev.to write-up of the reversal calls it a technical one ^[15]. Chaining one tool's result into another made the model read the full data and write it out again inside the conversation context, whether it ran to a hundred lines or a hundred thousand ^[7]. The top of that range is 1,000 times the bottom ^[1]. Every intermediate step cost tokens, and many servers returned plain text meant for a person, not structured data another program could parse ^[8]. One of Earendil's engineers wrote an entire post on why Pi would not support the protocol ^[3]. It now doubles as a problem statement for the fix.

The reversal, as Earendil tells it, started in its own harness. The company says the infrastructure MCP needed was useful regardless of MCP: lazy tool loading, system messages injected mid-conversation, and per-tool metadata for models running at different reasoning levels ^[9]. It redesigned Pi so those primitives are part of the core, and says MCP then fit without friction ^[10]. I think this is good engineering. Deferred loading and per-tool metadata go after the same context budget the composition complaint was about, and Earendil says they pay off beyond MCP ^[9][7].

Composition is handled one layer up. Codemode, as the write-up describes it, gives the client a JavaScript sandbox that combines several MCP calls without spending context on each intermediate step ^[11]. MCP still does discovery and invocation; the sandbox holds the logic that joins the calls ^[11]. The write-up says Codex, OpenAI's assistant, exposes an equivalent mechanism, and that several agent harnesses are moving toward letting the model write the combination logic instead of running a fixed plan of calls ^[12].

The same post says several agent tools went from distrust to adoption ^[13]. It names Pi as the case and Codex as the parallel, and does not name the others ^[12]. The evidence that critics are converging on MCP is one company's reversal, told partly through that company's blog ^[3]. In my view MCP is a sensible default for the announcement format, and a server costs a JSON file and an npx command to stand up ^[14]. What carries over from Pi is the order of work: Earendil changed the harness first and added the protocol second ^[10]. A client that connects MCP servers without deferred loading or code-based composition gets the version Earendil refused, with the token cost it described ^[7][8].