AI NewsDev toolsAnnouncement

Pi-Bolt packages the Pi coding agent as one native executable

Pi-Bolt packages the Pi coding agent as one native executable. The maintainers say it is 10 times faster to start and uses 4.5 times less memory than Pi on Node.

AI News

Editorial2 min read

LinkedInX
GitHub social card for opensec-git/Pi-Bolt

Image: GitHub

Why it mattersA coding agent that wakes in 41 ms stays open all day. Each prompt costs less, so a long session does not get slow. Node.js is not required.

A coding agent that pauses for half a second every time you type a command adds up over a day. Pi-Bolt is the Pi coding agent repackaged, with every function compiled to machine code when the executable is built. The maintainers released version 1.0.3 this week. The repository has 580 stars under the MIT licence.

What Pi-Bolt is

The project is a build of Pi, a terminal-based coding agent from Earendil Works. Pi normally runs on Node.js or Bun. Pi-Bolt is one native file. The maintainers list builds for Linux x86-64 and ARM64, macOS on Apple silicon, Windows x64 and Android in Termux. The maintainers say Pi-Bolt uses the same commands, sessions, settings and providers as Pi. They say every Pi extension works unchanged. The build in this release is Pi 1.4.0 inside a new runtime.

The numbers the project publishes

The maintainers publish side-by-side timings against Pi 1.4.0 on Node 24, pinned to one CPU on GitHub's Linux x86-64 runners. The project reports that Pi-Bolt is ready to type 41 ms after launch, against 423 ms for Pi on Node. It reports a 400-kilobyte file write through a tool call in 1.2 seconds, against 143.7 seconds. Streaming a long answer uses 2.3 seconds of CPU, against 18.6 seconds. Each prompt into a 40-prompt session uses 46 ms of CPU, against 352 ms. Peak memory after 40 prompts is 117 MB, against 521 MB. The project reports similar ratios on Linux ARM64, Apple silicon and Windows x64. These are the project's own measurements and no outside party has repeated them yet.

What release 1.0.3 changed

The release notes say an answer into a 4,000-entry session now costs 0.41 seconds of CPU on Linux x86-64. The number in 1.0.2 was 0.88 seconds. The maintainers say a frame no longer copies the whole transcript. A long streamed answer draws about 19 percent fewer frames. A 60,000-character answer uses about 25 percent fewer instructions. The release also fixes pi-bolt update when the GitHub API rate limit is reached, and extensions that use top-level await.

What is still Pi

The project does not change what Pi draws on screen or which requests it sends. It changes how fast Pi starts, how much CPU each turn costs and how much memory the process keeps. A session you leave open with 4,000 messages behaves close to a fresh one. Pi extensions and settings carry over.

Teams that already run Pi can switch with one install command. Some teams gave up on a terminal coding agent because the Node start-up delay felt slow. These numbers are a reason to try again. The cost is that the project is maintained by a third party, not by Earendil. Only the project itself has published the benchmarks so far.

Source

SourcePi-Bolt

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project