Pi 1.0 release notes report prompt cut from about 5,300 to 3,300 tokens

The comparison uses GPT-5.6 with default tools and Codemode. Pi finds MCP tools on demand rather than putting every tool definition in the prompt.

The comparison uses GPT-5.6 with default tools and Codemode. Pi finds MCP tools on demand rather than putting every tool definition in the prompt.
Pi 1.0 reduced a GPT-5.6 request's prompt from roughly 5,300 to 3,300 tokens when using the default tools and Codemode, according to release notes reported by The New Stack on October 5. The prompt is the text sent to the model; tokens are the units of text it processes. Pi is a coding agent, and Codemode is its sandbox for calling tools.
The release notes attribute the reduction to shortening Codemode's description, moving documentation for calling models out of the prompt and removing repeated tool definitions for tools that scripts could already access. This is a prompt-size comparison, not an independent test of speed, cost or task performance.
MCP support reversed an earlier position: Pi creator Mario Zechner had dismissed the protocol as unnecessary. The New Stack reports that he had measured Chrome DevTools MCP, a server for browser tools, at roughly 18,000 tokens before useful work began—about 9% of a 200,000-token context window. That window is the amount of information a model can use at once, so tool descriptions consumed space before the agent started its task.
Pi provides an agent harness: the software around a superintelligence (SI) model that connects it to tools and manages the conversation. The release adds Codemode, which Earendil describes as a sandbox within that harness to help agents issue tool calls—requests to use connected tools. Codemode supports MCP, decision models and image models.
By default, Pi puts a one-line description of each MCP server in the system prompt, which gives the model its operating instructions, rather than loading all the server's tool definitions. The agent uses Codemode to discover tools as needed and call them from JavaScript. Scripts can filter and combine results before returning useful output to the model. Developers can change this behavior with the toolExposure setting: even within one server, they can expose selected tools directly to the model, keep others behind Codemode or block them entirely.
Connecting a model to tools is different from coordinating an application that spans several conversations. Earendil says Pi Durable handles that second job: it orchestrates the conversations and manages storage using SQLite, a database engine, and JSONL, a format that stores JSON records one per line. That storage holds state—the information the application needs to keep as work proceeds.
Pi Durable does not replace the Pi coding agent. Earendil presents it as a framework for building long-running SI agent applications, coding agents included. The company says keeping the framework separate preserves Pi's minimalist design: developers can use Pi for model-and-tool conversations and the additional framework when an application needs to coordinate several conversations and retain their state.

The desktop-only tool combines Sidekick chat with direct editing and interactive previews, but does not initially support third-party themes.
2 min

AWS says the tool turns its best-practice guidance into application-specific recommendations and proposed fixes. Access requires an AWS support plan.
2 min

An interim report describes two different warning sequences: the following driver acknowledged a warning at a red signal, while an incorrect warning at a green signal had stopped the train ahead.