The MCP surface
The MCP surface
A fourth surface (after PHP, Artisan, HTTP API): expose the guardrails to AI clients through the Model Context Protocol via laravel/mcp. Default-OFF.
Enable
Install laravel/mcp
composer require laravel/mcpTurn it on
AI_GUARDRAILS_MCP_ENABLED=trueStart the local server
php artisan mcp:start ai-guardrails
When mcp.enabled is true and laravel/mcp is installed, the provider registers a local (stdio) server under the handle ai-guardrails. When the package is absent the registration gracefully no-ops.
The tools
| Tool | Does |
|---|---|
screen_prompt |
screen a prompt (Control B) → { blocked, rule_id, ruleset_version?, message? } |
sanitize_output |
sanitize an untrusted text blob (Control C) → { text } |
recent_injection_audit |
list recent audit attempts (read-only) |
flowchart LR
Client[AI / MCP client] -->|tools/call| Server[GuardrailsMcpServer]
Server --> T1[screen_prompt]
Server --> T2[sanitize_output]
Server --> T3[recent_injection_audit]
T1 & T2 & T3 --> Facade[AiGuardrails facade]
External-surface hardening
The MCP tools are treated as an externally-reachable surface:
- Input is length-capped (
sanitize_outputrejects oversized payloads) to prevent CPU/memory amplification. ruleset_versionis withheld on benign screens — exposing it on every call would let an adversarial client fingerprint the active ruleset.principal_idis omitted from the audit tool by default, opt-in viainclude_principal_ids, so internal user identifiers aren’t leaked.
The Laravel\Mcp\* reference is confined to src/Mcp (architecture test). Mount the server however your host runs MCP — laravel/mcp provides the transport; this package provides the tools.
Last updated:
Edit this page
Recent commits
