Install, update and verify

Pix has three install surfaces, and each is a separate step:

Surface	Installed by	Reaches
Hooks	vulnetix agent install	Claude Code and OpenAI Codex, the two hosts whose hook contract has been run against the CLI
Skills	vulnetix skills install, npx skills add, or the Claude Code plugin	Every agent in the table below, and more through npx skills
Subagents	The Claude Code plugin only	Claude Code

vulnetix agent install does not copy skills and does not write MCP configuration. It prints each host’s skills directory and the MCP URL so you can see where they go.

1. Install the CLI

Every hook and skill runs the Vulnetix CLI, which must be on the PATH of the process that starts your agent.

brew install Vulnetix/tap/vulnetix
# or
curl -fsSL https://cli.vulnetix.com/install.sh | sh

Scoop and Nix also work (scoop install vulnetix, nix profile install github:Vulnetix/cli). See Install the CLI for Windows and CI.

2. Wire the hooks

vulnetix agent install

It detects the agents on this machine, writes the hook entry into each host’s own file in its own format, keeps everything else in that file, and reports what changed. Running it again prints already configured.

Flag	Default	Effect
--agent <id>	every detected agent	Wire only this agent. Repeatable. An unknown id fails with unknown agent "<id>"; run 'vulnetix agent hosts' for the list.
--dry-run	off	Print what would change; write nothing.
--no-hooks	off	Skip hook configuration and only report skills directories and MCP.

Each host gets one entry per event, all running the same command, vulnetix agent hook, with a 30-second timeout:

Event	Matcher	Answers
PreToolUse	Bash|Edit|Write|MultiEdit (Codex: *)	Dependency guard, change guard
SessionStart	none	Session context
UserPromptSubmit	none	Session context

The entry is recognised by its command, so a re-run replaces this CLI’s entry in place and never touches your other hooks. The Claude Code plugin ships the same three entries in vulnetix/hooks/hooks.json, so a marketplace install has the hooks without running agent install.

At the end it says which credential the guard will use: Authenticated., or that it is running on the shared Community tier with the link to get a free key.

3. Install the skills

vulnetix skills install

It uses the first tool it finds, in this order:

Tool found	Runs
npx	npx skills add --all -y Vulnetix/pix-ai-coding-assistant
claude (no npx)	claude plugin marketplace add Vulnetix/pix-ai-coding-assistant, then claude plugin install vulnetix@vulnetix-plugins
gh (neither)	gh skills add Vulnetix/claude-code-plugin <skill> --agent <agent> -f --scope user for each detected agent and each skill in the latest release

--agent <id> and --skill <name> target one agent or one skill, and then every tool that is present is used (npx skills add Vulnetix/pix-ai-coding-assistant --agent <id> for npx).

Either registry tool works directly too, because the skills live at skills/ in the repository root:

npx skills add Vulnetix/pix-ai-coding-assistant

Claude Code plugin

Inside Claude Code:

/plugin marketplace add Vulnetix/pix-ai-coding-assistant
/plugin install vulnetix@vulnetix-plugins

The marketplace is vulnetix-plugins and the plugin is vulnetix (version 2.0.0). The plugin carries the 18 skills, the 5 subagents and the 3 hook entries. Skills are invoked as /vulnetix:<skill>, for example /vulnetix:repo-impact CVE-2021-44228.

What each agent gets

From vulnetix agent hosts, the table the installer acts on. vulnetix agent hosts -o json prints it as JSON.

Id	Agent	User skills directory	Project skills directory	Hooks	MCP
claude-code	Claude Code	~/.claude/skills	.claude/skills	yes, ~/.claude/settings.json	yes
codex	OpenAI Codex	~/.agents/skills	.agents/skills	yes, ~/.codex/hooks.json	yes
cursor	Cursor	~/.agents/skills, ~/.cursor/skills	.agents/skills, .cursor/skills	not verified	yes
gemini-cli	Gemini CLI	~/.agents/skills, ~/.gemini/skills	.agents/skills, .gemini/skills	not verified	yes
github-copilot	GitHub Copilot	~/.copilot/skills	.github/skills	not verified	yes
opencode	opencode	~/.config/opencode/skills	.opencode/skills	not verified	yes
pi	Pi	~/.agents/skills, ~/.pi/agent/skills	.agents/skills, .pi/skills	not verified	yes
windsurf	Windsurf	~/.codeium/windsurf/skills	.windsurf/skills	not verified	yes
zed	Zed	~/.agents/skills	.agents/skills	not verified	yes
amp	Amp	~/.config/agents/skills	.agents/skills	not verified	yes

An agent counts as installed when its configuration directory exists (~/.claude, ~/.codex, ~/.cursor and so on). “Not verified” means the host documents hooks but this CLI has not been run against them, so agent install wires nothing and says hooks not configured: <agent>'s hook contract has not been verified against this CLI.

For MCP, point the agent at https://mcp.vulnetix.com/mcp. See MCP server.

Update

Installed with	Update with
vulnetix skills install	vulnetix skills update (runs npx skills update -y, else gh skills update --all, else claude plugin marketplace update)
Claude Code plugin	/plugin update vulnetix
Hooks	Nothing to update: the hook is the CLI. Upgrade the CLI (brew upgrade vulnetix).

No updater deletes a skill that a release removed. vulnetix skills update and vulnetix skills check list leftovers, and vulnetix skills prune removes them:

vulnetix skills prune          # print what would be removed
vulnetix skills prune --yes    # remove it

Prune reads ~/.agents/.skill-lock.json and only touches entries the lock attributes to a Vulnetix source. Name skills (vulnetix skills prune fix verify-fix --yes) to remove those even if the release still ships them.

Verify

Check	Command	Expect
CLI on the agent’s PATH	command -v vulnetix && vulnetix --version	A path and a version
Which credential	vulnetix auth status	Your org, or the Community tier
Hosts and surfaces	vulnetix agent hosts	INSTALLED yes for your agent, HOOKS yes for Claude Code and Codex
Hooks wired	vulnetix agent install --dry-run	hooks already configured
Skills present	vulnetix skills check	The Vulnetix skills in each directory, and any missing or retired ones
Hook answers	echo '{"hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"npm i left-pad"},"cwd":"'"$PWD"'"}' | vulnetix agent hook	Usually nothing: silence is a pass. See Hooks.

In Claude Code, /hooks lists three vulnetix agent hook entries (PreToolUse, SessionStart, UserPromptSubmit) and /plugin lists vulnetix as enabled.

Credentials

None are required. The CLI looks for one in this order and uses the first it finds: VULNETIX_API_TOKEN; VULNETIX_API_KEY with VULNETIX_ORG_ID; VVD_ORG with VVD_SECRET; ./.vulnetix/credentials.json; ~/.vulnetix/credentials.json; the OS keyring; .netrc. With none, it runs on the shared Community credentials. A free Community key with its own quota is at https://www.vulnetix.com/resolve/register, then:

vulnetix auth login

It prints a URL and a code and waits until you approve in the browser. An agent cannot approve it for you. See Authentication.