Getting Started

nervepack is a context hub you run across your machines — skills, memory, tools, and workflows your AI coding assistant reads every session. This is the first-time install.

Prerequisite: an agentic host — one that can read/write files and run shell commands (Claude Code, Cursor, Codex, Goose, …). A plain chat UI can only consume the knowledge as context; it can’t self-wire.

Quick start

The same four steps on Linux, macOS, and Windows. The host self-wires — you don’t hand-edit hooks or settings; the onboard does it and proves it with the doctor.

1. Get the code.

git clone https://github.com/pat-browne/nervepack ~/Code/nervepack

Git is the only prerequisite for this step. On Windows, install Git for Windows first — it supplies the bash that runs every nervepack hook and cron.

2. Install the toolchain (gh, jq, node, python, go, build tools):

OS	Command
Linux	python3 ~/Code/nervepack/engine/nervepack_engine/cli.py setup install-apt-baseline
macOS	python3 ~/Code/nervepack/engine/nervepack_engine/cli.py setup install-brew-baseline
Windows	No installer step — install Node 20, Python 3, gh, and jq by hand (Git for Windows already covers bash).

The engine runs on Bash + Python 3 — the baseline above installs Python (apt on Linux, uv python install on macOS; install it by hand on Windows), so every engine script works out of the box. On a Claude host, finish with python3 ~/Code/nervepack/engine/nervepack_engine/cli.py setup install-claude-plugins.

3. Authenticate GitHub (so the maintenance jobs can push your content):

gh auth login    # GitHub.com · HTTPS · login with a browser

4. Onboard the host. Open your agent in ~/Code/nervepack and say:

onboard nervepack

(On a fresh box there is no slash command yet — /np-onboard is itself a nervepack skill, so it only exists after onboarding links the skills. Saying “onboard nervepack” works from zero because the agent auto-loads this repo’s CLAUDE.md, which routes it to the contract. Non-interactive alternative: python3 ~/Code/nervepack/engine/nervepack_engine/cli.py onboard runs the same wiring directly. Once onboarded, /np-onboard is available for re-runs.)

Your agent reads the tool-neutral contract, links the skills, installs the session hooks and scheduler, writes ~/.config/nervepack/adapter.json, and runs the doctor until every MUST capability is green. Verify any time:

python3 ~/Code/nervepack/engine/nervepack_engine/cli.py doctor   # per-capability PASS/MISSING; non-zero on a real gap

That’s the whole install. What differs by OS is handled for you:

	macOS	Windows
Toolchain	cli.py setup install-brew-baseline	manual (Node · Python · gh · jq)
Scheduler backend	launchd LaunchAgents	Windows Task Scheduler
Hook execution	native bash	commands auto-wrapped through Git-bash

On a non-Claude host, there’s no CLAUDE.md auto-load or slash command — point your agent at the same contract directly: ../engine/onboard/ONBOARD.md.

Going deeper

Everything above gets you running. These child pages cover the optional, heavier setup:

• Content & team overlays — point the engine at your own private skills/memory repo (and an optional shared team overlay). Recommended; without it the engine falls back to its own root and you have nowhere personal to grow.
• Scheduled maintenance agents — the crons/routines that promote memory, compact skills, and lint. Optional; nervepack works without them.
• Onboarding contract — the tool-neutral capability contract for wiring any agentic host by hand.
• Bootstrapping over MCP — after the one git clone, point any MCP client at the server and a single nervepack_onboard call wires the rest.
• Architecture · Features — how the engine, layers, and toggles fit together.

Sanity-checking paths

The engine and your content overlay live in separate repos, so a script named in a skill can drift. This catches a stale path before you chase a dead command:

python3 ~/Code/nervepack/engine/setup/np-path-check.py ~/Code/nervepack ~/Code/nervepack-content

A clean run prints all setup/onboard path references resolve ✓; any hit names the file, the line, and the path to fix. The same check runs in CI, so the engine’s own docs stay honest.