Blitz
The zero-gimmick, high-performance Go coding agent.
Blitz reads your workspace, makes the change, checks it, and gets out of the way: no persona, no filler, no chatter. It is written in Go on Google’s Agent Development Kit and runs Gemini, Claude, OpenAI-compatible or Ollama models.
blitz "add unit tests for user_service.go" # one prompt, then exit
blz -d ~/src/api # a session in another project
| Program | |
|---|---|
blitz, the CLI |
An interactive session in the terminal, or a one-shot run for scripts and CI. A single static binary that starts in about 20 ms. |
blitzd, the service |
A per-user service holding every workspace you open, reached over a Unix socket only you can open. It runs scheduled workers and lets the CLI and the desktop app share sessions. |
| Blitz, the desktop app | A window onto the service, laid out like an IDE: the workspace’s files, an editor, and the agent’s chat. |
All three run the same engine, so a session started in the terminal can be picked up in the desktop app. The architecture explains how they fit together.
- Compiled Go. Instant start, low memory, no runtime to install. See performance.
- Safe by default. Commands run in an OS sandbox, edits need your approval unless you’ve allowed them, and every turn can be undone. See the safety model.
- Works with what you have.
AGENTS.md,CLAUDE.mdandGEMINI.mdload as project instructions; MCP servers, hooks, custom commands and Agent Skills all work. - Scriptable. JSON and streaming JSON output, limits on turns, cost and time, and distinct exit codes.
- Open. Apache License 2.0. Every requirement is written down in the specifications.
- New to Blitz: Getting started, then a tutorial from an empty folder to working software.
- Using it: the products and the guide.
- Changing it: architecture, shared packages and development.