Install

Requirements

Needed for	What
Running Gru	Claude Code (the skill also runs in Codex; the tether MCP step is Claude-only)
Cheaper minions	tether (installed for you via uv) and/or opencode
Sandboxing (Linux, WSL2)	bwrap (bubblewrap): sudo apt install bubblewrap. Without a sandbox Gru refuses to run minions.
Sandboxing (macOS, native Windows)	Docker (Docker Desktop, Colima or Podman providing the docker CLI). Untested there so far. Also optional on Linux for QA with real services.
Delivery	git and gh for worktrees, branches and PRs
NVIDIA models	NVIDIA_API_KEY in your environment (~/.profile). Never printed or committed.

One command

curl -fsSL https://raw.githubusercontent.com/igormcsouza/gru/main/install.sh | bash

Choose the host with ... | bash -s -- --target claude|codex|both|other. It clones to ~/.gru (override with --dir or GRU_HOME) and symlinks:

• the skill (SKILL.md plus its ref/ files) into ~/.claude/skills/gru (Codex: ~/.agents/skills/gru)
• the dev-team agents (team-leader, planner, swe, qa, reviewer) into ~/.claude/agents/

For Claude Code it also runs uv tool install --upgrade "tether-sdk[mcp]", adds an nvidia provider to ~/.tether/providers.yaml if missing and registers the tether MCP server at user scope. Failures here warn and never abort.

It is idempotent and never overwrites a real file: a SKIPPED line means something real already uses that name, so move it away and rerun. Rerun to update.

From a clone

git clone https://github.com/igormcsouza/gru.git ~/Projects/gru
~/Projects/gru/install.sh --target claude

Verify

tether --version
ls -l ~/.claude/skills/gru ~/.claude/agents

Start a new session (skills and agents load at session start), then run /gru in a project: the first run interviews you, and /gru selftest checks the harnesses (once per machine and after upgrading one).