Pi drops its no-MCP stance and adds a tool-call sandbox

Pi's site used to declare that the harness does not support MCP, and its authors said the same on podcasts and in a November 2025 post by Mario Zechner. Earendil now says MCP is part of Pi's core, after being available as a community extension (pi-mcp-adapter). The post gives two reasons: MCP has changed, and the changes Pi needed for it (tools that can be deferred or made available only to the sandbox) were useful on their own.

The other piece is Codemode, which the post describes as a sandbox that "runs where the harness runs." The agent uses JavaScript there to order and combine tool calls, and its state is kept in the session transcript instead of the file system. Pi loads it automatically when MCP is configured. Earendil's stated view is that MCP "should be much closer to OpenAPI with intelligent tool discovery," and it says many MCP servers are still built for harnesses that dump every tool into context.

The post includes an example session, described as a condensed replay, in which the agent pulls 167 open issues through a Linear MCP server and rates each thread with a classifier model. It reports 11 flagged as mildly frustrated. The post shows no measurements of token use or success rate, and its claim that this avoids wasting context is Earendil's own.

Why it matters: For anyone choosing or building a harness, this is one concrete design for keeping MCP tool definitions out of the model's context: expose them to a sandbox and let the agent script the calls. The post offers an example, not a benchmark, so test it against your own tool set before relying on it.