Arthur AI

Deploy the integrations/arthur-ai FastAPI wrapper on any public HTTPS host. The gateway calls it at llm_input / llm_output via the Custom Guardrail contract; the wrapper forwards traffic to Arthur GenAI Engine and returns verdict JSON on HTTP 200.

What is Arthur AI?

Arthur GenAI Engine is a SaaS platform for validating LLM prompts and completions. The wrapper calls POST /api/v2/validate and maps rule results to pass/block — it does not embed policy logic beyond the checks you configure. Arthur is validate-only. Use Operation: Validate for both input and output rails. It reports failures but does not redact or rewrite text. Default checks when config is {}:

Input: PromptInjectionRule + ToxicityRule
Output: ToxicityRule

How it works

The gateway POSTs an OpenAI-shaped requestBody (input) or requestBody + responseBody (output) to your wrapper URL.
The wrapper extracts user/assistant text and calls Arthur with your ARTHUR_API_KEY.
The wrapper returns HTTP 200 with a policy outcome in the body. Infrastructure failures return HTTP 5xx.

TF customer's app ─► TF gateway ─► this wrapper ─► engine.platform.arthur.ai/api/v2/validate

Response contract

HTTP	Body	Meaning
`200`	`{"verdict": true}`	Allow
`200`	`{"verdict": false, "message": "..."}`	Block (policy)
`5xx`	error JSON	Wrapper or Arthur failure

Policy blocks use 2xx + verdict: false, not HTTP 4xx. See Custom guardrail response contract.

Wrapper endpoints

Path	Target
`/validate-input`	Request (input)
`/validate-output`	Response (output)

GET /health — health check. GET /debug/loaded-config — bearer-gated diagnostics. All POST routes expect Authorization: Bearer <WRAPPER_API_KEY>.

Prerequisites

Arthur API key from the Arthur platform.
Public HTTPS URL for the deployed wrapper.
WRAPPER_API_KEY — shared secret the gateway sends as Authorization: Bearer ….

Setup

Clone and configure

git clone https://github.com/truefoundry/integrations-custom-guardrails
cd integrations-custom-guardrails/integrations/arthur-ai
cp .env.example .env

.env

ARTHUR_API_KEY=<from https://platform.arthur.ai>
WRAPPER_API_KEY=<generate: python -c "import secrets; print(secrets.token_urlsafe(32))">

Deploy the wrapper

TrueFoundry:

pip install -U truefoundry
tfy login
python deploy.py --wait

Set TFY_WORKSPACE_FQN, TFY_PUBLIC_HOST, TFY_PUBLIC_PATH, and secret FQNs in .env. Create secrets arthur-api-key and wrapper-api-key under group arthur-guardrails-tfy in Platform → Secrets.Local:

python3 -m venv .venv
.venv/bin/pip install -r requirements-dev.txt
.venv/bin/uvicorn main:app --reload --port 8000

AI Gateway → Guardrails → + Add New Guardrails Group → type Custom.

Group name: arthur-ai
Add two configs — input and output.

Input validate example:

Field	Value
Name	`arthur-input-validate`
Operation	Validate
Target	Request
Enforcing Strategy	Enforce But Ignore On Error
URL	`https://<host>/validate-input`
Headers	`Authorization` → `Bearer <WRAPPER_API_KEY>`; `Content-Type` → `application/json`
Config	`{}`

TrueFoundry custom guardrail form: Validate, Request target, /validate-input URL, Authorization Bearer header — Arthur AI Custom Guardrail configuration (input validate)

Output validate:

Field	Value
Name	`arthur-output-validate`
Operation	Validate
Target	Response
URL	`https://<host>/validate-output`

Auth Data → Custom Bearer Auth works the same as Headers if you prefer not to set headers manually.

Attach to traffic

Model pin: AI Gateway → Models → <model> → Guardrails → attach group arthur-ai.Per request — X-TFY-GUARDRAILS header, selector format <group>/<config-name>:

{
  "llm_input_guardrails": ["arthur-ai/arthur-input-validate"],
  "llm_output_guardrails": ["arthur-ai/arthur-output-validate"]
}

Custom config (optional)

Override defaults by setting config.checks in the dashboard:

{
  "checks": [
    {"name": "prompt-injection-check", "type": "PromptInjectionRule", "apply_to_prompt": true, "apply_to_response": false},
    {"name": "toxicity-check", "type": "ToxicityRule", "apply_to_prompt": true, "apply_to_response": false, "config": {"threshold": 0.5}}
  ],
  "fail_closed_on_unavailable": false
}

Key	Purpose
`credentials.apiKey`	Override `ARTHUR_API_KEY` env var
`api_base`	Override Arthur API host (default `https://engine.platform.arthur.ai`)
`timeout`	Request timeout in seconds (default 30)
`context` / `grounding_context`	Grounding text for hallucination checks
`fail_closed_on_unavailable`	Block when Arthur returns Skipped/Unavailable (default `false`)

Troubleshooting

Symptom	Likely cause
`401` from wrapper	`WRAPPER_API_KEY` on the service does not match the dashboard Bearer token
Gateway allows despite `verdict: false`	Tenant gateway not honoring verdict-on-200; set Enforce or upgrade gateway
Arthur Skipped/Unavailable but traffic allowed	Default behavior; set `fail_closed_on_unavailable: true` in config
Wrong checks running	Curl `/debug/loaded-config` with Bearer auth to inspect loaded config

Reference

Item	Value
Source repo	`truefoundry/integrations-custom-guardrails/integrations/arthur-ai`
Arthur platform	platform.arthur.ai
Arthur API	`POST https://engine.platform.arthur.ai/api/v2/validate`
Selector	`arthur-ai/<config-name>`

Get Started

LLM Gateway

MCP Registry and Gateway

Skills Registry

Prompt Registry

Guardrails and Security

Observability

Deployment

Admin Guide

Chat

Messages

Embeddings

Rerank

Responses

Image

Audio

Batch

Files

Fine-tuning

Moderations

Models

What is Arthur AI?

How it works

Response contract

Wrapper endpoints

Prerequisites

Setup

Custom config (optional)

Troubleshooting

Reference

​What is Arthur AI?

​How it works

​Response contract

​Wrapper endpoints

​Prerequisites

​Setup

​Custom config (optional)

​Troubleshooting

​Reference

What is Arthur AI?

How it works

Response contract

Wrapper endpoints

Prerequisites

Setup

Custom config (optional)

Troubleshooting

Reference