magi-code.
Built for real work.

I moved from Cursor to Claude Code, then OpenCode and Pi. Each taught me something useful—and each left part of the experience outside my control.

Customizing an existing agent often meant working around a large fixed prompt, fighting an interface decision, or carrying a difficult fork as upstream changed. I wanted fine-grained control over the prompt, the tools, the flow of work, and every small interaction in the terminal.

magi-code became the place to be unapologetically particular about all of it. It also gave me a practical way to learn Rust: not through a toy, but through a tool I use to build itself.

The point is not another prompt box with tool output below it. magi-code is designed around the shape of serious, evolving repository work.

• Primary-agent-led flow. One agent holds the objective, makes decisions, and remains accountable for the result.
• Subagents with a real job. Parallel workers have focused identities, bounded context, visible progress, and structured results.
• Tools tuned for context. Reading, searching, editing, browsing, and delegation are designed to return what the agent needs without flooding its attention.
• Mission Control. The TUI adds an activity tree, full-history navigation, and a working view of what the agent system is doing—not just a prettier transcript.
• History without amnesia. Compaction keeps current context useful while rotated session logs remain available when earlier detail matters.

Performance matters because long-running agent work should not become disposable simply because the transcript is large.

One continuous parent session used to build magi-unreal-axi ran from August 3–31, 2026. It remained practical to open and navigate with the complete activity tree available.

magi-code runs locally and connects to supported model providers. Install it with Cargo, authenticate, then launch it from the repository you want to work in.

cargo install magi-code --locked
cd your-repository
magi-code

View package and release history ↗