coding-agent

Terminal coding assistant built on mu-core + mu-harness + mu-tui. Uses mu-local-provider (llama-swap) for inference and mu-ai-tools for filesystem + shell access. Sessions persist as JSONL under $XDG_DATA_HOME/mu/sessions/; plugins, skills, agents, permissions load from $XDG_CONFIG_HOME/mu/.

Run from source with pnpm dev (tsx). mu install <npm:spec | ./path.ts> adds a plugin; mu -c resumes the most recent session.

Project README

mu

Minimal terminal AI coding agent for local models (llama-swap / llama.cpp).

Install

curl -fsSL https://raw.githubusercontent.com/gaetan-puleo/mu-ai/main/install.sh | sh

Installs the standalone mu binary to ~/.local/bin (no Node needed). On Windows, grab mu-windows-x64.exe from the releases page.

Configure

~/.config/mu/config.json:

{
  "kind": "llama-swap",
  "baseUrl": "http://localhost:8080"
}

Run

mu        # start
mu -c     # resume the last session

Type to chat. @file/@agent to mention, !cmd to run a shell command, / for commands (/new, /sessions, /model, /thinking, /expand, /voice, /call, /context-export, /quit). /voice dictates one clip into the composer and /call is hands-free realtime dictation — both need a microphone recorder (ffmpeg / arecord / parecord) and an audio-capable model, or voiceModel set to one. Tab cycles the agent, Esc Esc cancels, Ctrl+T toggles the theme.

Packages

core/            # agent loop: messages, tools, provider, createAgent
harness/         # sessions, agents, skills, sub-agents, permissions, scheduler, plugins, chat TUI
tui/             # terminal UI engine (parser, layout, components)
tools/           # read, write, edit, bash, list
local-provider/  # llama-swap / llama.cpp provider
webfetch/        # URL → markdown tool
coding-agent/    # the `mu` CLI

Develop

pnpm install       # workspace deps
pnpm dev           # run the mu CLI from source (tsx)
pnpm test          # vitest
pnpm check         # type-check (tsc)
pnpm -r build      # build all packages (tsup)

License

MIT