Pi, the minimalist coding agent, reverses course on MCP

Four tools: `read`, `write`, `edit` and `bash`. That was the minimal inventory Pi launched with, the terminal coding agent that Mario Zechner, creator of the libGDX game framework, develops as open source in the pi-mono repository. Among its deliberate omissions was one that stood out at the height of the protocol's growth: no MCP.

That refusal has expired. According to The Register, Pi has done a 180 and is adding support for the Model Context Protocol. The change carries more weight than it seems, because Zechner had been one of the clearest voices against using MCP in coding agents, to the point of titling one of his posts “What if you don't need MCP at all?”.

Where Pi came from

Pi was born as a reaction to agents that kept piling on features. Its pitch was a short system prompt, no built-in subagents or plan mode, support for several model providers and a TypeScript extension system so each user could add only what they needed. That lightweight base is also what OpenClaw uses as its core, so Pi's design decisions reach further than its own community.

The case against MCP was mostly about context cost. Popular servers such as Playwright's or Chrome DevTools' load twenty or more tool definitions into every session, and those tokens are spent before the agent does anything useful. The alternative Zechner advocated was writing small command-line tools paired with a README that the agent reads only when it needs them, an idea very close to Claude Code's skills.

Why the reversal makes sense

Two things have changed. The first is the weight of the standard. MCP, created by Anthropic and donated in December 2025 to the Agentic AI Foundation under the Linux Foundation umbrella, has become the usual way to expose services to a model. GitHub, Linear, Notion and Sentry publish their own MCP servers, and an agent that does not speak the protocol forces its user to rebuild every integration by hand.

The second is that the context problem now has partial answers. In November 2025, Anthropic explained in its post on code execution with MCP that an agent treating servers as APIs called from code, instead of loading every definition at once, can go from 150,000 to 2,000 tokens in an example workflow, a 98.7% reduction. Claude Code, for its part, already loads MCP tools on demand when their definitions take up too much context. Zechner's objection still holds, but it no longer forces a choice between standards and efficiency.

It remains to be seen how the support fits with Pi's extension philosophy and how much context it consumes by default. For a project that made restraint its signature, how the feature is switched on matters as much as the feature itself. That is where the line will fall between a pragmatic concession and abandoning the original idea.

Who it is useful for

Pi users who already maintained MCP servers for Claude Code or Claude Desktop. Since this is a standard protocol, those servers should work without being rewritten, even if configuration differs from one tool to another.
Teams that publish MCP servers, who gain one more client with a technical user base that pays close attention to token consumption.
* Anyone designing their own agents, because the Pi case shows that even the most minimalist positions end up negotiating with interoperability.

We have seen the same dilemma in client projects: every connected MCP server makes an integration easier and, at the same time, takes up space the model needs to work. We think this is a reasonable reversal; what is worth keeping from Zechner's original stance is the discipline of making every server justify the tokens it costs.