Installation
Rocky is a single binary with no external dependencies. You download it and it just works.
Requirements
- A terminal
- Claude Code (recommended) or OpenCode β Rockyβs primary mode runs as skills inside your coding agent, using its context window for extraction. Rocky itself doesnβt call an LLM in this flow, so no API key is needed.
For workflows outside an agent session β running rocky from a terminal, or a plain post-commit hook without the Claude Code skills β you can optionally configure a standalone LLM provider. See Alternative workflows at the bottom of this page.
Option 1: Install with Cargo (recommended)
If you have Rust installed:
cargo install --git https://github.com/NVME-git/rocky
This downloads, compiles, and installs Rocky in one step. It will be available as rocky from anywhere on your system.
Donβt have Rust? Install it from rustup.rs β it takes about 2 minutes.
Option 2: Download a prebuilt binary
Go to the Releases page and download the binary for your operating system.
Linux / macOS:
chmod +x rocky
mv rocky ~/.local/bin/rocky
Windows: Move rocky.exe somewhere and add that folder to your PATH.
Verify the installation
rocky stats
You should see a banner and an empty PKG:
Rocky Β· Personal Knowledge Graph
ββββββββββββββββββββββββββββββββββββββ
Total topics: 0
Known: 0
Fading: 0
Gaps/weak: 0
PKG is empty. Let us begin science, question?
Set up Claude Code (recommended)
This is the canonical Rocky workflow. One command wires up the four skills, the prompt-logging hook, and the git queue:
rocky install claude-all
claude-all drops these four skills into ~/.claude/skills/:
/rocky-checkpointβ end-of-session: read recent diffs + transcript, extract topics with generic question banks (no repo-specific identifiers), merge into the PKG (using the global dedup list so the same idea across projects becomes one node with multiplerepos[])./rocky-quizβ Socratic review inside the Claude session, picking from the canonical question bank./rocky-backfillβ one-shot seeder for projects that already had history when Rocky was installed./rocky-promptiqβ re-evaluates your recent prompts and produces a PromptIQ score with feedback.
It also enables prompt logging so the skills can see your session transcript, and rewires the git post-commit hook so commits queue diffs (which /rocky-checkpoint drains) instead of running an LLM directly.
For OpenCode or another compatible coding agent, point it at ~/.claude/skills/ (or copy the skill files into the agentβs equivalent directory) β the skills are agent-generic markdown.
Granular install options if you want pieces individually:
rocky install skills # just the skills, into ~/.claude/skills/
rocky install claude # just the prompt-logging hook
Youβre ready. Head to Quick Start for your first session.
Alternative workflows (standalone CLI without an agent)
These are for when you want Rocky to run outside an agent session β say, a plain
Plain git post-commit hook
Skip the queue mode and have rocky diff run directly after each commit:
rocky install # or: rocky install git
Requires a provider configured below.
Provider option A: Ollama (free, runs locally)
The recommended standalone provider. Local, free, no API key.
- Install Ollama from ollama.com
- Pull a model:
ollama pull qwen2.5-coder:7b
- Create
~/.config/rocky/config.toml:
[llm]
provider = "ollama"
model = "qwen2.5-coder:7b" # or "llama3.1:8b" for general-purpose
See the Configuration page for model recommendations by GPU VRAM.
Provider option B: Anthropic (Claude API)
Use this if you donβt want to run a local model and youβd rather pay per-token.
- Get an API key from console.anthropic.com
- Set it in your environment:
# Add to ~/.bashrc or ~/.zshrc
export ANTHROPIC_API_KEY=sk-ant-your-key-here
Or create a .env file in your project directory.
Legacy: per-turn Stop hook
rocky install stop
Auto-extraction on every Claude turn via Ollama. Not wired by claude-all β the per-turn latency was a real complaint β but available as opt-in.
Coming from a Rocky version that wrote
~/.rocky/.rocky.toml? It auto-migrates on first run.