llm-setup

Lean Linux installer for Simon Willison's llm CLI, Anthropic's Claude Code, and a curated set of Claude Code skills.

Runs as a regular user on Debian / Kali (no WSL ties, no system-level hardening, no GUI integrations). Self-updates from git on every run.

Quick Start

git clone https://github.com/c0ffee0wl/llm-setup ~/llm-setup
cd ~/llm-setup
./linux/setup.sh

That installs llm with a trimmed plugin set, Claude Code, and copies the skills into ~/.claude/skills/. Provider configuration (Azure / Gemini keys, default model) is the user's responsibility — see below.

Provider setup

Setup seeds Azure model YAML templates into ~/.config/io.datasette.llm/ on first run. Configure your provider manually:

Azure OpenAI:

# 1. Replace __AZURE_API_BASE__ with your resource URL
#    e.g. https://YOUR-RESOURCE.cognitiveservices.azure.com/openai/v1/
nano ~/.config/io.datasette.llm/extra-openai-models.yaml
nano ~/.config/io.datasette.llm/azure-embeddings-models.yaml

# 2. Set your API key
llm keys set azure

# 3. Set the default model
llm models default azure/gpt-4.1-mini

Google Gemini:

llm keys set gemini
llm models default gemini-2.5-flash

Anthropic, OpenRouter, etc.: likewise, llm keys set <provider> then llm models default <model>.

Re-running ./linux/setup.sh will never overwrite an existing YAML in ~/.config/io.datasette.llm/, so your edits persist.

Setup Modes

Command	What it does
./linux/setup.sh	Install missing tools, upgrade already-installed ones, seed Azure YAML templates if absent
./linux/setup.sh --skip-skills	Same, but skip the skills sync (statusline still installed)
./linux/setup.sh --clear-cache	Purge package-manager caches (npm, go, pip, pipx, cargo, uv) and exit
./linux/setup.sh --uninstall	Remove what this script installed (prompts per group; keeps user data)
./linux/setup.sh --uninstall --dry-run	Preview the uninstall — list every action, change nothing
./linux/setup.sh --uninstall --force	Uninstall everything without per-group prompts

The install/upgrade run is idempotent and non-interactive: re-running is safe; the script never prompts. There is no --upgrade flag — installed components (llm + plugins, Claude Code, gitingest) are upgraded automatically on every run. llm is fully reinstalled only when its plugin list changes (tracked by a fingerprint); otherwise it gets a fast uv tool upgrade. imagemage is install-if-missing only (no per-run Go rebuild). Every normal run also purges package-manager caches at the end (same as a standalone --clear-cache).

If claude is actively running, setup refuses to proceed (Phase 0 self-update, claude update, and --uninstall can swap or remove the binary under a live session). Stop your Claude session, then re-run. --clear-cache is exempt — it changes no binaries and works during a session.

Uninstall

--uninstall removes only what this script installs, one group at a time (answer Y/n per group, or pass --force to remove all without asking; --dry-run previews):

• llm + all its plugins (the llm uv tool), together with its install-state files (llm-install-fingerprint, uv-tool-packages.json) — bundled so that keeping llm keeps its plugin manifest intact — and gitingest
• the Claude Code binary (~/.local/bin/claude) and ~/.local/bin/{imagemage,blaude}
• this repo's skills under ~/.claude/skills/ (local and external ones listed in skills/external-skills.yaml) and ~/.claude/statusline.sh
• the AppArmor bwrap profile (/etc/apparmor.d/bwrap, via sudo — unloaded from the kernel, then removed)

Preserved: your API keys, default model, and seeded Azure YAML in ~/.config/io.datasette.llm/; ~/.config/uv/uv.toml; a customised ~/.claude/settings.json (removed only if byte-identical to the fresh-install template); system apt packages (git curl jq bubblewrap) and runtimes (uv, pipx, go).

What's installed

• llm CLI (from the c0ffee0wl/llm fork) with: provider plugins (llm-gemini, llm-vertex, llm-anthropic, llm-openrouter), llm-cmd, llm-git-commit, llm-jq, llm-templates-fabric, fragment loaders (llm-fragments-github, llm-fragments-pdf, llm-fragments-site-text, llm-fragments-dir, llm-fragments-youtube-transcript), llm-sort, llm-classify, and llm-uv-tool (plugin persistence across uv tool upgrade).
• Claude Code via the official claude.ai/install.sh installer.
• Claude Code skills copied from this repo's skills/ to ~/.claude/skills/. External skills listed in skills/external-skills.yaml are refreshed on every run.
• Claude Code statusline at ~/.claude/statusline.sh (sourced from /opt/claude-litellm's richer LiteLLM-aware variant; degrades gracefully without LiteLLM). ~/.claude/settings.json is written only on first install — never clobbered.
• gitingest (uv tool) — turns a git repo into LLM-friendly text.
• imagemage (Go binary at ~/.local/bin/imagemage) — Gemini image-generation CLI used by the image-generation skill. Needs Go ≥ 1.22 (installed from apt if missing); silently skipped if apt can't supply a recent enough Go.
• blaude (script at ~/.local/bin/blaude) — a bubblewrap sandbox wrapper for Claude Code, fetched from c0ffee0wl/blaude. The bubblewrap package is installed from apt; on Kali/Ubuntu 24.04+ an AppArmor profile for bwrap is also written (/etc/apparmor.d/bwrap, via sudo) so unprivileged user namespaces work. The companion osc52-clipboard script is not installed.

Layout

llm-setup/
├── README.md                # this file
├── CLAUDE.md                # agent guidance
├── PLAN.md                  # design rationale
├── skills/                  # Claude Code skills (platform-agnostic)
└── linux/
    ├── setup.sh             # entry point (6 phases)
    ├── common.sh            # helper library
    ├── configs/             # Azure model YAML templates
    └── scripts/statusline.sh