The Pulse
Pi 1.0 Cuts Codemode Prompt Use by About 40%
Pi released version 1.0 of its open-source terminal coding agent on October 1, 2026. The update reduces codemode prompt tokens and adds image generation, while Pi’s documentation warns that its tools run with the permissions of the agent pr

AI.info Team · Written with AI ·
Pi 1.0 adds tighter MCP authentication and more ways to connect tools, but Pi’s documentation carries a blunt caveat: its tools and extensions run with the permissions of the Pi process, and project trust controls do not sandbox them. The release, published October 1, puts that tension alongside a practical efficiency claim: codemode uses about 40% fewer prompt tokens.
The update moves the terminal interface to fullscreen by default, adds image generation inside codemode scripts and streamlines sign-in for Radius and Anthropic. Those changes give developers more ways to extend and operate the agent, while leaving responsibility for the permissions it inherits with the user.
Pi says codemode uses fewer prompt tokens
Pi’s release notes describe the reduction both as “about 40% fewer prompt tokens” and through a specific example: with default tools and codemode active, a GPT-5.6 request shrinks from about 5,300 tokens to 3,300. The roughly 2,000-token reduction comes from shorter tool descriptions and system-prompt guidance, with the codemode reference carrying details the model can consult when needed.
Codemode lets the model write JavaScript that calls Pi’s tools and non-LLM models. Scripts can run calls in parallel and filter large results before returning output to the model. Pi also says errors now offer recovery guidance, such as suggesting a similarly named tool when a script refers to one that does not exist.
Fullscreen becomes the default
Pi’s terminal interface now starts in fullscreen mode. Users who want the terminal’s ordinary scrollback can set tuiMode to "regular" or launch Pi with --tui-mode regular, according to the release notes.
The version also adds models.generateImages() to codemode scripts, using credentials already associated with the session. Generated images count toward session cost, the release notes say. A separate usability change lets users complete Anthropic sign-in by copying a code when the browser is running on another machine.
MCP sign-in gets more controls
For Model Context Protocol servers, Pi adds an option to specify an OAuth authorization-server metadata URL when a server advertises incorrect metadata or none. The release also describes RFC 9207 issuer checks, credentials stored per server, and sign-in that preserves previously granted scopes when a server requests additional permissions.
Those controls address authentication boundaries, but they do not change the permission model for tools. Pi’s documentation says tools and extensions operate with the permissions of the Pi process; project trust determines which project resources load, not whether tool calls are isolated. Users handling unfamiliar repositories, extensions or unattended automation should account for that limitation.
Version 1.0 follows a rapid feature run
Pi’s changelog lists version 0.99.2 on September 30, one day before 1.0.0. That earlier release expanded MCP support and changed when servers connect; the 1.0 release builds on the same work with more compact codemode descriptions and additional authentication controls.
The project publishes the release through npm and GitHub, and its documentation identifies Pi as MIT-licensed. Version 1.0 is available now, with fullscreen as the default and the regular terminal mode still available through a setting or command-line option.