Arthur

Arthur is an AI observability, evaluation and governance platform that monitors ML models, LLM applications and AI agents in production. It is built for enterprise AI, platform and security teams that need guardrails against hallucination, PII leakage and prompt injection, plus continuous evals and an inventory of every agent running across the company.

Arthur is an AI observability, evaluation and governance platform from Arthur AI, a New York company founded in 2019 and led by CEO and co-founder Adam Wenchel, which has raised more than $60M, including a $15M Series A led by Index Ventures in December 2020 and a $42M Series B co-led by Acrew Capital and Greycroft in September 2022. It began as a machine-learning monitoring tool that tracked accuracy drift, bias and explainability for tabular, NLP and computer-vision models, and has since been rebuilt around generative AI and agents. Its core, the Arthur Engine, was open-sourced under the MIT licence in April 2025 and combines a GenAI Engine (real-time guardrails for PII, hallucination, prompt injection and toxic language, with hallucination checked claim by claim against the supplied context), an ML Engine for drift and accuracy metrics, OpenTelemetry/OpenInference trace collection, prompt management, and continuous evaluations that score groundedness, tool selection, relevance and PII leakage. The commercial Arthur Platform adds dashboards, custom alerting and webhooks, experiments that replay gold-standard datasets against new agent versions, discovery of unregistered shadow agents across endpoints, cloud and private data centres, behavioural runtime security, and token-cost tracking with rerouting to cheaper models, streaming findings to Splunk, Datadog, Elastic and CrowdStrike Falcon. It runs as SaaS, in a dedicated or managed VPC, or self-hosted via Docker Compose. Named users include Axios, The Philadelphia Inquirer, Expel and Upsolve, which used Arthur evals to catch a GPT-5 regression before customers saw it. Its earlier open-source LLM evaluation tool, Arthur Bench, was archived in September 2026. Arthur competes with Arize AI, Fiddler AI, Galileo and LangSmith.

Category

Governance & Security

Pricing

Freemium, Subscription, Contact for pricing

Target Market

CIOs, CTOs, CISOs, ML Engineers, Enterprise Developers, Data Scientists

Deployment

Cloud-first, Self-hosted, Open-source

Founded

2019

Headquarters

New York, USA

• ✓
  Real-time LLM guardrails

  Detects and blocks PII, hallucinations, prompt injection and toxic language in model inputs and outputs, reducing the risk of harmful or non-compliant responses reaching users.

• ✓
  Claim-level hallucination detection

  Checks groundedness claim by claim against message history or retrieved context, so RAG and agent answers unsupported by source data are flagged.

• ✓
  Continuous evals and experiments

  Scores agents on groundedness, tool selection, relevance and PII leakage, and replays gold-standard datasets against new versions to surface regressions before release.

• ✓
  OpenTelemetry agent tracing

  Collects OpenInference/OpenTelemetry traces with token counts, latencies and retrieval performance, giving engineers step-level visibility into multi-step agent behaviour.

• ✓
  Agent discovery and governance

  Finds registered and shadow agents running on endpoints, cloud and private data centres and streams findings to SOC tools such as Splunk and CrowdStrike Falcon.

• ✓
  ML model monitoring

  Tracks drift, accuracy metrics and model comparisons for traditional machine-learning models, so classic and generative AI are monitored in one place.

• ✓
  Cost control

  Tracks token spend per application and can reroute traffic to cheaper models, helping teams keep agent operating costs under control as usage scales.

• •
  Catching model-upgrade regressions

  A product team replays gold-standard datasets through its agent after switching LLMs; Upsolve used this to catch a GPT-5 regression before customers noticed.

• •
  Guarding customer-facing chatbots

  An enterprise places Arthur guardrails in front of a support assistant to block prompt injection and redact PII in real time before responses are returned.

• •
  Governing shadow AI agents

  A security team discovers unregistered agents running across endpoints and cloud accounts and routes the findings into its existing SIEM for investigation and policy enforcement.

• •
  Monitoring production ML model drift

  A financial-services data science team monitors credit or fraud models for drift and accuracy decay and receives alerts before business metrics degrade.

• •
  Self-hosted RAG evaluation

  A regulated company deploys the open-source Arthur Engine with Docker inside its own environment to evaluate RAG pipeline groundedness without sending data to SaaS.

List pricing is published: a Free tier (4 use cases, 7-day retention) and Premium at $60/month, metered by use cases, jobs, spans, inferences and evals rather than seats (seats are unlimited). SSO, SLAs, a BAA, VPC deployment and unlimited retention require custom-priced Enterprise. The Arthur Engine itself is free and open source under the MIT licence.

This page was written from 10 sources, 6 on domains other than arthur.ai.

1. 1.arthur.ai — arthur.aivendor
2. 2.arthur.ai — pricingvendor
3. 3.github.com — arthur engine
4. 4.github.com — bench
5. 5.techcrunch.com — arthur ai snags 15m series a to grow machine learning monito
6. 6.prnewswire.com — arthur raises 42m in series b funding as ai adoption soars 3
7. 7.news.ycombinator.com — item
8. 8.arthur.ai — how upsolve built trusted agentic ai with arthurvendor
9. 9.arthur.ai — arthur achieves soc 2 r type ii certification compliancevendor
10. 10.docs.arthur.ai — docs.arthur.ai