AgentGuards plugin for Claude Code — self-hosted

For a self-hosted AgentGuards appliance. If you're on the hosted product at agentguards.co, install agentguards-claude instead — it bundles the MCP server this variant deliberately does not.

Why a separate plugin

Claude Code's plugin .mcp.json cannot substitute an environment variable into its url field — it has to be a fixed string, baked in when the plugin is published. agentguards-claude's .mcp.json therefore points at https://prod.agentguards.co, and there is no way to override that per installation. Shipping one plugin for both audiences meant a self-hosted operator's MCP tool calls (check_input, authorize_action, …) would go to AgentGuards' hosted service instead of their own appliance, with no error and no way to tell — the opposite of what a self-hosted deployment is for.

This variant ships hooks only, with no MCP server at all. Hooks read AGENTGUARDS_URL from the environment as a subprocess, correctly per installation, so they aren't affected by that limitation. They are also the actual enforcement mechanism — check_input/authorize_action in the other variant are a cooperative convenience the model can choose to call; the hooks run regardless. Nothing here is a reduced version of the guardrail, only a narrower plugin.

Install

/plugin marketplace add alelaguard/agentguards-plugins
/plugin install agentguards-claude-selfhosted@agentguards

Then point it at your appliance. Unlike the hosted plugin, there is no default — this is deliberate, so it can never silently talk to the wrong instance:

export AGENTGUARDS_URL=https://<your-appliance>
export AGENTGUARDS_API_KEY=ag_your_token_here

Add both lines to your shell profile (~/.bashrc, ~/.zshrc, …) and restart Claude Code. Or run /agentguards:setup and it will walk you through it, including the appliance's first-boot self-signed certificate if that applies.

Alternative: npm install @agentguardsco/claude-selfhosted-plugin. Fetches these files for programmatic use — it does not register with Claude Code on its own; use /plugin install above for that.

Commands

• /agentguards:setup — configure the appliance URL and API key, and verify.
• /agentguards:status — report reachability and whether the key is accepted.

Configuration

| Variable | Required | Default | Purpose | |---|---|---|---| | AGENTGUARDS_URL | yes | — none | Your appliance's own address. No fallback, by design. | | AGENTGUARDS_API_KEY | yes | — | Your ag_ token, generated on your appliance's own /admin/ui/keys. | | AGENTGUARDS_CA_BUNDLE | no | — | Pin the appliance's certificate while it's still self-signed. Verification stays on. | | AGENTGUARDS_TLS_NO_VERIFY | no | false | Skip certificate verification. Private-network evaluation only. | | AGENTGUARDS_FAIL_OPEN | no | false | Hooks fail closed by default. Set true to allow on error. |

How it works

The hooks call your appliance's REST API on every prompt, before every Bash command, and after every web fetch — blocking or redacting when it flags a risk. There is no cooperative MCP layer in this variant; see the bundled guardrails skill for why that's not a gap.

Learn more at https://agentguards.co.