guardrails（guardrails 监控告警）

Routing note: For ambiguous user intents, use the shared clarification templates in references/intent-clarification.md.

Guardrails

Configure content safety guardrails for TrueFoundry AI Gateway. Guardrails add safety controls to LLM inputs/outputs and MCP tool invocations.

When to Use

Set up guardrail providers, create guardrail rules, or manage content safety policies for AI Gateway endpoints. This includes PII filtering, content moderation, prompt injection detection, secret detection, and custom validation rules.

Deploying a Custom Guardrails Server

When the user asks to deploy a guardrails server or run guardrails as a deployed service, start from the official template so the server adheres to the gateway's input/output formats:

1. Clone the default repo: truefoundry/custom-guardrails-template
2. Build on top of it — Add or adjust custom rules, providers, or config within the template structure; do not build from scratch.
3. Deploy — Use the deploy skill to deploy the resulting service (Dockerfile or build from source as in the template).

This keeps guardrail servers compatible with TrueFoundry AI Gateway expectations.

Overview

Guardrails require a two-step setup:

1. Guardrail Config Group — Register guardrail provider integrations (credentials and configuration)
2. Gateway Guardrails Config — Create rules that reference those providers and attach them to a gateway

Step 1: Create Guardrail Config Group

A guardrail config group holds integration credentials for one or more guardrail providers. See references/guardrail-providers.md for all supported providers.

When using direct API, set TFY_API_SH to the full path of this skill's scripts/tfy-api.sh. See references/tfy-api-setup.md for paths per agent.

List Existing Config Groups

Via Tool Call

tfy_guardrail_config_groups_list()

Via Direct API

TFY_API_SH=~/.claude/skills/truefoundry-guardrails/scripts/tfy-api.sh

$TFY_API_SH GET '/api/svc/v1/provider-accounts?type=guardrail-config-group'

Create Config Group

Via Tool Call

tfy_guardrail_config_groups_create(payload={"name": "my-guardrails", "type": "provider-account/guardrail-config-group", "integrations": [...]})

Note: Requires human approval (HITL) via tool call.

Via Direct API

$TFY_API_SH POST /api/svc/v1/provider-accounts '{
  "name": "my-guardrails",
  "type": "provider-account/guardrail-config-group",
  "integrations": [
    {
      "type": "integration/guardrail/tfy-pii",
      "config": {}
    },
    {
      "type": "integration/guardrail/tfy-content-moderation",
      "config": {}
    }
  ]
}'

Each integration has a type (from the providers reference) and a config object with provider-specific fields. Some providers (like tfy-pii, tfy-content-moderation) require no config. Others (like aws-bedrock, azure-content-safety) need cloud credentials.

Security: Guardrail providers with external endpoint_url fields (e.g., custom, opa, fiddler, palo-alto-prisma-airs) route request data to third-party services. Verify that all external endpoints are trusted and controlled by your organization before registering them. Prefer TrueFoundry built-in providers (tfy-pii, tfy-content-moderation, tfy-prompt-injection) when possible.

Presenting Config Groups

Guardrail Config Groups:
| Name             | ID       | Integrations |
|------------------|----------|--------------|
| my-guardrails    | pa-abc   | 3            |
| prod-safety      | pa-def   | 5            |

Step 2: Create Gateway Guardrails Config

Gateway guardrails config defines rules that control which guardrails apply to which models, users, and tools.

Get Existing Guardrails Config

Via Tool Call

tfy_gateway_guardrails_list()

Via Direct API

$TFY_API_SH GET /api/svc/v1/gateway-guardrails-configs

Create Guardrails Config

Via Tool Call

tfy_gateway_guardrails_create(payload={"name": "production-guardrails", "type": "gateway-guardrails-config", "gateway_ref": "GATEWAY_FQN", "rules": [...]})

Note: Requires human approval (HITL) via tool call.

Via Direct API

$TFY_API_SH POST /api/svc/v1/gateway-guardrails-configs '{
  "name": "production-guardrails",
  "type": "gateway-guardrails-config",
  "gateway_ref": "GATEWAY_FQN",
  "rules": [
    {
      "id": "pii-filter-all-models",
      "when": {
        "target_conditions": {
          "models": ["*"],
          "mcp_servers": [],
          "tools": []
        },
        "subject_conditions": {
          "users": ["*"],
          "teams": []
        }
      },
      "llm_input_guardrails": [
        {
          "provider_ref": "provider-account-id:integration/guardrail/tfy-pii",
          "operation": "validate",
          "enforcing_strategy": "enforce",
          "priority": 1
        }
      ],
      "llm_output_guardrails": [
        {
          "provider_ref": "provider-account-id:integration/guardrail/tfy-pii",
          "operation": "validate",
          "enforcing_strategy": "enforce",
          "priority": 1
        }
      ],
      "mcp_tool_pre_invoke_guardrails": [],
      "mcp_tool_post_invoke_guardrails": []
    }
  ]
}'

Update Existing Guardrails Config

Via Direct API

$TFY_API_SH PUT /api/svc/v1/gateway-guardrails-configs/GUARDRAILS_CONFIG_ID '{
  "name": "production-guardrails",
  "type": "gateway-guardrails-config",
  "gateway_ref": "GATEWAY_FQN",
  "rules": [...]
}'

Rule Structure

Each rule contains:

• id — Unique identifier for the rule
• when — Conditions controlling when the rule applies:

- target_conditions.models — Model name patterns (use ["*"] for all) - target_conditions.mcp_servers — MCP server names to target - target_conditions.tools — Specific tool names to target - subject_conditions.users — User patterns (use ["*"] for all) - subject_conditions.teams — Team names

• llm_input_guardrails — Applied to LLM request inputs
• llm_output_guardrails — Applied to LLM response outputs
• mcp_tool_pre_invoke_guardrails — Applied before MCP tool execution
• mcp_tool_post_invoke_guardrails — Applied after MCP tool execution

Guardrail Reference Fields

Each guardrail entry in a rule has:

• provider_ref — Format: <provider-account-id>:integration/guardrail/<provider-type>
• operation — validate (check and block) or mutate (modify content, e.g., redact PII)
• enforcing_strategy — How violations are handled:

- enforce — Block the request on violation - audit — Log the violation but allow the request - enforce_but_ignore_on_error — Enforce if guardrail succeeds, allow if guardrail errors

• priority — Integer for ordering when multiple mutate guardrails apply (lower runs first)

Common Patterns

PII Detection on All Models

# Step 1: Create config group with tfy-pii
$TFY_API_SH POST /api/svc/v1/provider-accounts '{
  "name": "pii-guardrails",
  "type": "provider-account/guardrail-config-group",
  "integrations": [
    {"type": "integration/guardrail/tfy-pii", "config": {}}
  ]
}'

# Step 2: Create rule targeting all models
# Use the provider account ID from step 1 response in provider_ref

Content Moderation with Audit Mode

Use "enforcing_strategy": "audit" to log violations without blocking — useful for monitoring before enforcement.

MCP Tool Guardrails

Target specific MCP tools with mcp_tool_pre_invoke_guardrails to validate inputs before tool execution, or mcp_tool_post_invoke_guardrails to scan tool outputs.

Model-Specific Rules

Use target_conditions.models to apply guardrails only to specific models:

"when": {
  "target_conditions": {
    "models": ["openai/gpt-4*", "anthropic/claude-*"],
    "mcp_servers": [],
    "tools": []
  }
}

Exempt Specific Users

Combine broad model targeting with specific user conditions to exempt admin users:

"subject_conditions": {
  "users": ["user1@example.com", "user2@example.com"],
  "teams": ["engineering"]
}

Finding the Gateway Reference

The gateway_ref is the fully qualified name (FQN) of your AI Gateway deployment. Use the ai-gateway skill to list gateways and get the FQN.

<success_criteria>

Success Criteria

• The user can list existing guardrail config groups
• The user can create a new guardrail config group with the desired provider integrations
• The user can create or update gateway guardrails config with rules
• Rules correctly target the intended models, users, and tools
• The agent has confirmed create/update operations before executing
• Provider references correctly link to the config group integrations

</success_criteria>

Composability

• Preflight: Use status skill to verify credentials before configuring guardrails
• Requires ai-gateway: Get the gateway FQN for gateway_ref
• Requires access-control: For subject exemptions in rules (users, teams)
• References mcp-servers: For MCP tool guardrail targets
• Provider reference: See references/guardrail-providers.md for all 23 supported providers

Error Handling

Config Group Not Found

Provider account not found. List config groups first to find the correct ID.

Invalid Provider Type

Unknown guardrail integration type. Check references/guardrail-providers.md for valid types.

Gateway Not Found

Gateway reference not found. Use the ai-gateway skill to list available gateways.

Duplicate Rule ID

Rule ID already exists in this config. Use a unique ID for each rule.

Missing Provider Credentials

Integration config missing required fields. Check the provider reference for required config.

Permission Denied

Cannot manage guardrails. Check your API key permissions.