PI-Desktop Is Betting Your Coding Agent Should Never Phone Home
PI-Desktop jumped 636 stars in a day to 2,250, which for an Electron app with a Rust core is a lot of people quietly agreeing with its premise. The premise: your coding agent should run on your machine, with your keys, and no mandatory cloud relay in the middle.
It is a desktop workspace built on Electron plus a Rust host core, wrapping the pi agent harness. Bring your own model, OpenAI or Anthropic or a local endpoint or anything OpenAI-compatible. Three modes: Agent runs autonomously, Plan waits for approval on each step, Goal chases an outcome. Multi-project and multi-session management with pinning and branching, plus built-in review panels for diffs, command output, and app previews. It extends through Skills, MCP servers, subagents, and user-installable plugins, and it will import sessions from other coding agent tools.
The local-first part is not decoration. No mandatory relay, no required account, credentials stored in the OS keychain rather than on someone's server. Given that the last two weeks produced a session-token infostealer campaign and a threat report about traffic being silently forwarded to a third-party model, "your keys never leave the keychain" is starting to read as a security feature rather than a philosophy.
LGPL-3.0, version 0.14.x, early preview, macOS Windows and Linux builds at https://github.com/vastsa/PI-Desktop
The honest caveat is that this category is crowded and half of it is a wrapper around someone else's harness with a nicer diff view. What makes PI-Desktop worth a look is the plugin surface plus the session-import path, which together say the ambition is to be where you run every agent rather than the place you run one. Whether a two-thousand-star early preview can hold that position against OpenAI shipping a free managed loop the same week is a genuinely open question.
Related reading: https://clauday.com/article/c7742540-f008-4ab3-bbce-137ca81bcc96
← Back to all articles
It is a desktop workspace built on Electron plus a Rust host core, wrapping the pi agent harness. Bring your own model, OpenAI or Anthropic or a local endpoint or anything OpenAI-compatible. Three modes: Agent runs autonomously, Plan waits for approval on each step, Goal chases an outcome. Multi-project and multi-session management with pinning and branching, plus built-in review panels for diffs, command output, and app previews. It extends through Skills, MCP servers, subagents, and user-installable plugins, and it will import sessions from other coding agent tools.
The local-first part is not decoration. No mandatory relay, no required account, credentials stored in the OS keychain rather than on someone's server. Given that the last two weeks produced a session-token infostealer campaign and a threat report about traffic being silently forwarded to a third-party model, "your keys never leave the keychain" is starting to read as a security feature rather than a philosophy.
LGPL-3.0, version 0.14.x, early preview, macOS Windows and Linux builds at https://github.com/vastsa/PI-Desktop
The honest caveat is that this category is crowded and half of it is a wrapper around someone else's harness with a nicer diff view. What makes PI-Desktop worth a look is the plugin surface plus the session-import path, which together say the ambition is to be where you run every agent rather than the place you run one. Whether a two-thousand-star early preview can hold that position against OpenAI shipping a free managed loop the same week is a genuinely open question.
Related reading: https://clauday.com/article/c7742540-f008-4ab3-bbce-137ca81bcc96
Comments