Pi integrates MCP with Codemode to reduce token usage in AI agents

Initial Token Consumption Challenge

Mario Zechner, creator of the Pi coding agent, previously avoided integrating the MCP protocol due to its high token cost. Measurements showed that Chrome DevTools MCP alone consumed approximately 18,000 tokens, or 9% of a 200,000-token context window, before any useful operations. Playwright MCP also required about 13,700 tokens for its 21 tools.

The Codemode Solution

Pi now incorporates MCP, but it does not expose all server tools to the model upfront. Instead, Pi utilizes Codemode to identify and call tools from JavaScript only when needed, returning only the relevant output. This method prevents the model from being burdened with large tool definitions in the initial prompt.

Improved Composability and Efficiency

Zechner also noted that traditional MCP servers lacked composability, requiring results to pass through the agent's context for persistence or combination. His earlier workaround involved using Bash and scripts, which required only a 225-token README and allowed direct piping or saving of output. The current Codemode approach in Pi addresses these composability issues by managing tool interactions more efficiently.

Earendil's Rationale for Integration

Earendil, which acquired Pi, decided to integrate MCP into the core of Pi 1.0. The company stated that MCP had matured sufficiently, and the architectural changes required for its integration were found to be broadly beneficial. This aligns with Pi's existing use of Codemode with Jev, applying the same pattern to manage interactions between the model and its tools.