Pi 1.0 tops Hacker News as the agent that rejected MCP ships native support

What happened

Pi, a minimalist terminal coding agent, took the top spot on Hacker News with its 1.0 release, collecting over 1,200 points according to a dev.to writeup. The interest was not just about the ranking: the project had spent more than a year publicly criticising the Model Context Protocol (MCP), and this release ships native MCP support anyway. A companion release, Pi Durable, also reached the site's top 10.

What Pi is

Pi is a deliberately spare coding-agent harness that runs in the terminal, works with models from every major provider and is MIT licensed. According to the dev.to post, Earendil — the company behind it, founded by Flask creator Armin Ronacher and backed by Accel and Balderton — claims hundreds of thousands of weekly users. Earendil took the project over in spring 2026; original creator Mario Zechner remains a shareholder and still steers the technical direction.

The post describes a stated philosophy of waiting until a technology proves itself before adopting it, weighing its function against the complexity it adds, and notes that the 'things we said no to' list in the release notes is longer than the feature list. That is the backdrop against which the MCP decision landed.

What 1.0 ships

Codemode is the headline feature: native MCP support plus a way to plug non-LLM components such as classifiers and image models into the agent. Around it sit deferred tool loading, so tools are not pushed into context up front; cache warming for Anthropic models; mid-conversation system messages, which let prompts and tools change partway through a transcript; virtual model extensions; and a redesigned terminal UI that defaults to full screen.

The MCP reversal, explained

Pi's old complaint was context cost. Tool schemas load up front, at a commonly cited 550 to 1,400 tokens per tool definition, and Zechner's go-to example was Perplexity, where three MCP servers reportedly consumed 143,000 tokens of a 200,000-token window.

Codemode does not load the schemas and live with the cost. The model instead gets a JavaScript sandbox, compiled to WASM and running inside the harness. It reads tool documentation, writes code that calls tools as if using an SDK, chains calls in parallel and filters the output, so only the distilled result re-enters the context. The announcement's demo combines the Linear MCP server with a sentiment classifier to surface frustrated commenters across 167 open issues, with all 167 processed in parallel and only the final answer reaching the model's context.

The post lists three stated reasons for the change: a July 2026 MCP spec revision that made the protocol more stateless, the fact that the sandbox was needed anyway for non-LLM models, and the view that a year of public criticism pushed the protocol to improve. Its author's conclusion is that much of what was called a protocol problem was really a harness problem.

Pi Durable: the second release

Shipped the same day, Pi Durable is an explicitly experimental framework for long-running, crash-resistant agents reachable from multiple surfaces such as a terminal or Slack. Per the post, it is around 15,000 lines of TypeScript excluding tests — estimated at 150,000 tokens to read with GPT and 250,000 with Claude, small enough that the whole codebase fits in a context window. It offers three storage backends (in-memory, SQLite and JSONL), with the SQLite one avoiding Node APIs so it runs on Bun or Cloudflare Durable Objects.

Every step is a durable checkpoint, so a dead process resumes rather than restarts, and conversations can fork from any point in a transcript. Tools declare a replay policy: those marked safe re-run after a crash, while unmarked side effects such as deploys or payments do not, and the model is instead told the call was interrupted. Interrupted model requests are resent behind a requestId intended to give exactly-once submission, and approval decisions are stored in a first-write-wins memo so a crash after a human approves does not trigger a second prompt. The framework is TypeScript-only for now, and a single process owns a given storage at a time, with API changes expected.

Why it matters

Two things in this release point at where agent tooling is heading. First, MCP looks increasingly settled as the wire protocol between agents and tools, but the interface presented to the model is changing shape — from a pile of tool schemas in context to a programmable environment. If Codemode's approach spreads, tool output stops being a tax on the context window, and other harnesses will likely converge on something similar. Second, durability is emerging as the gap between agent demos and production systems; checkpointing combined with explicit replay policies addresses crash behaviour that most frameworks skip entirely. Pi's reversal carries weight of its own: a project that built credibility by refusing things said yes, and its reasoning for doing so is documented in public.