alpine-code (CLI)

A terminal coding agent built on alpineagents. If you use Claude Code, Codex, opencode or pi, you already know how to use it.

  ▀▀   ▀▀    alpine-code v0.1.0
  ██   ██    model anthropic/claude-sonnet-5
▄█████████▄  cwd   ~/projects/my-app
███▀█▀█▀███  /help for commands · ctrl+d to exit
 ▀▀█████▀▀

› fix the failing test in tests/test_api.py
◆ read(tests/test_api.py)
  └  Read 84 lines
◆ bash(uv run pytest tests/test_api.py -x)
  └  FAILED tests/test_api.py::test_create - KeyError: 'id'

Install

uv tool install alpine-code     # or: pipx install alpine-code

Configure a model

export ALPINE_MODEL=anthropic/claude-sonnet-5   # uses ANTHROPIC_API_KEY
export ALPINE_MODEL=openai/gpt-5                 # uses OPENAI_API_KEY
export ALPINE_MODEL=ollama/qwen3-coder           # local Ollama

# Any OpenAI-compatible server (OpenRouter, vLLM, LM Studio, ...)
export ALPINE_MODEL=<model> ALPINE_BASE_URL=https://.../v1 ALPINE_API_KEY=...

Or save connections in ~/.alpine-code/config.toml (the desktop app writes the same file):

default_model = "anthropic/claude-sonnet-5"   # <connection>/<model>

[connections.anthropic]
provider = "anthropic"          # anthropic, openai, google, openrouter, zai-coding-plan, kimi-for-coding, minimax-coding-plan

[connections.local]
base_url = "http://localhost:11434/v1"

A connection's key comes from its provider's variable (ANTHROPIC_API_KEY, ZAI_API_KEY...) or from ~/.alpine-code/auth.json, where the app saves keys (readable only by you). ALPINE_CODE_HOME moves the folder.

Use

alpine                        # interactive
alpine "explain this repo"    # interactive, starting with a message
alpine -p "summarize README"  # print the answer and exit (scripts, CI)
git diff | alpine -p "review this diff"

Commands: /help, /clear, /compact, /model [name], /mode [mode], /cost, /exit.

Project instructions: AGENTS.md files from the git root down to the current directory are added to the system prompt (CLAUDE.md is read where there is no AGENTS.md). A global one can go in ~/.alpine-code/AGENTS.md.

Permissions (every rule is skipped in yolo):

• Reading and searching inside the working directory never asks. Paths outside it (symlinks resolved) ask, and so does reading .env / .env.* files (.env.example is fine).
• Editing asks unless the mode is accept edits.
• bash asks. "Don't ask again" remembers command prefixes for the session — git status, uv run pytest, npm run dev — parsed with tree-sitter, so git status; rm -rf x or git status $(curl …) still ask, as do commands that write files through >. Commands that can run anything (python, sudo, bash -c, xargs, …) are never remembered.
• bash also checks the paths it names — arguments, --flag=/path values and redirection targets, following cd — with the same rules as the file tools. Paths known only at run time ($HOME/.ssh, cd "$(…)") ask every time. Remembering an outside directory for bash means edit access to it, so that is only offered under your home directory or a temp directory; system directories, your home directory itself and / always ask.
• With -p, calls that would ask are declined.

Tools

Deliberately few. More will come as they prove necessary.

glob and grep run ripgrep, so they respect .gitignore. The rg on your PATH is used when there is one; otherwise a pinned release is downloaded once (checksum-verified) into ~/.cache/alpine-code/ripgrep/.