Arthur

Arthur is a platform for finding, observing, and governing AI agents across enterprise systems. It best suits organizations that need operational visibility and enforceable controls across several deployment environments. The combination is unusually broad, but smaller teams should weigh the Free and Premium quotas carefully before standardizing on it.

Overview

Arthur’s discovery layer looks beyond a single agent framework: sensors monitor OpenTelemetry streams, MCP servers, network traffic, cloud APIs, and employee endpoints. That makes it a strong candidate when agents are distributed across departments or infrastructure. It is less compelling for a team operating one tightly controlled stack that needs only basic tracing.

Key features

Tracing covers agent steps, tool calls, retrieved context, costs, latency, and behavior over time. Drift monitoring, model-performance metrics, and bias monitoring add ongoing oversight rather than limiting analysis to individual runs. Slack and webhook alerts provide practical ways to route events into existing response processes.

Governance uses machine-checkable policies and real-time guardrail verdicts for prompt injection, PII exposure, restricted models, and off-policy tool use. Those controls are suited to teams that must intervene while agents operate, not merely review logs afterward. Arthur’s open-source Evals Engine adds built-in PII, sensitive-data, custom-LLM, and regex rules, with self-serve deployment for evaluation workflows.

Documented integrations cover OpenAI, Anthropic, LangChain, LiteLLM, CrewAI, LlamaIndex, OpenAI Agents, Google ADK, AWS Bedrock, and Mastra. Connectors include BigQuery, Google Cloud Storage, Amazon S3, and Arthur Shield instances, giving data teams established destinations and sources to incorporate into monitoring.

Pricing

The Free plan costs $0/mo and is suitable for a limited pilot: up to 4 use cases, 1 organization, 1 workspace, 2 projects, 7-day retention, 5k jobs, 300k spans, 12k inferences, and 3k evals. Seats are unlimited, but the short retention and small project allowance make it unsuitable for broad, long-lived production coverage.

Premium costs $60/mo and expands capacity to 100 use cases, 10 projects, 30-day retention, 20k jobs, 1.2M spans, 100k inferences, and 75k evals, while retaining one organization and one workspace. It fits a growing team that needs materially more room without enterprise procurement, although the fixed quotas still require volume planning.

Enterprise uses custom pricing and removes the principal scale constraints with unlimited organizations, workspaces, projects, and data retention, plus custom jobs, spans, inferences, and evals. It adds dedicated and managed VPC options, SSO, SLAs, and a BAA, along with a dedicated customer success manager and advanced monitoring. This is the appropriate tier for regulated or large deployments requiring contractual support and infrastructure control.

Platforms

Arthur is offered through an API, web interface, and self-hosted deployment. Hosted SaaS, on-premises, and hybrid choices accommodate different operational policies; the hybrid evals engine runs beside customer workloads so sensitive data remains in the environment. RBAC, SSO, and deployed engineering support are stated for every deployment option.

Who it's for

Choose Arthur when agents span multiple enterprise surfaces and you need discovery, trace-level diagnostics, runtime guardrails, and policy enforcement in one system. It is particularly well matched to security, platform, and compliance teams managing sensitive data or restricted models. A small project with modest telemetry volume may find the Free plan adequate, while organizations needing durable history or many projects should budget for Premium or Enterprise.

Pros and cons

Pros

• Broad discovery: Sensors reach telemetry, infrastructure, cloud, and endpoint sources, helping uncover agents outside a single framework.
• Runtime governance: Real-time verdicts target prompt injection, PII, restricted models, and off-policy tools.
• Deployment flexibility: SaaS, on-premises, and hybrid operation support different data-control requirements.
• Deep integration list: Major agent frameworks and cloud data connectors reduce the need to isolate monitoring from existing systems.

Cons

• Free-tier constraints: Seven-day retention, two projects, and defined job, span, inference, and eval quotas limit sustained use.
• Premium remains bounded: One organization and workspace, plus monthly volume caps, may not suit complex enterprises.
• Enterprise requires a quote: Larger deployments must go through custom commercial planning.

Alternatives

Evidently AI is the better pick when a fully open-source Apache 2.0 framework and self-managed deployment are priorities.

Deepchecks fits teams wanting a freemium entry tier with three seats, one AI application, and prompt-based metrics.

NannyML suits buyers focused on self-managed monitoring, with an open-source plan and a $399/month Starter tier.

Opik is an alternative when locally runnable open-source observability and evaluation are central requirements.

Amazon SageMaker Autopilot is for users who prefer Amazon’s pay-as-you-go API and web service.

Arize AX offers a free SaaS tier with 25k trace spans per month and 15-day retention.

Fiddler AI is an alternative for teams prioritizing real-time guardrails and trace-based plans.

Galileo suits organizations seeking a Pro plan with 50,000 traces per month and dedicated Slack support.

For broader comparisons, see Machine Learning Model Monitoring Software and Model Monitoring Software.

Verdict

Arthur is a strong choice for enterprises that need to discover agents across heterogeneous systems, inspect their behavior, and enforce runtime policies while retaining deployment control. Its main reason to look elsewhere is quota economics: the Free tier is deliberately limited, Premium is still capped, and the scale and contractual features that larger organizations may need require custom Enterprise pricing.