Connect your AI Agents to Arize in minutes

What is an Arize MCP?

It's an MCP server that gives your agents access to Arize's AI observability and evaluation platform through tools. Your agents can invoke these tools to list and create monitors, manage prompt versions, query datasets, build annotation queues, and run evaluations, and more.

Arize offers an official MCP server, but you can use one from a third-party platform, like Merge Agent Handler.

How can I use the Arize MCP server?

The use cases naturally depend on the agent you've built, but here are a few common ones:

• Prompt regression detection: When a new prompt version is deployed in your LLM application, an agent retrieves the latest evaluation scores from Arize for that version and compares them against the prior version, posting a Slack alert if performance drops below a defined threshold before the change reaches production users

• Monitor alert routing: When a model monitor in Arize fires, an agent reads the monitor details, identifies the affected model and metric, and creates a Linear issue assigned to the owning ML engineer with the full threshold violation details attached

• Dataset preparation before evaluation runs: When a new evaluation is kicked off, an agent adds examples to the relevant Arize dataset from an external data store, ensuring the evaluation set is current and complete before the evaluators execute

• Annotation queue creation for human review: When model performance drops below a defined threshold, an agent creates an annotation queue in Arize and routes it to the appropriate review team, ensuring low-quality outputs get human review rather than going unexamined

What are popular tools for Arize's MCP server?

Here are some of the most commonly used tools:

• list_monitors: returns all monitors configured for a model in Arize, including their status, metric type, and threshold settings. Useful for agents that need a snapshot of active monitoring coverage before deciding whether to create a new alert or update an existing one

• list_models: retrieves all models registered in an Arize space. Call this when an agent needs to identify which model to scope a monitor, evaluation, or dataset to before taking further action

• list_prompt_versions: returns all versions of a prompt along with their metadata and labels. Good for workflows that compare prompt performance across versions or need to identify the current production-tagged version before running an evaluation

• create_prompt: registers a new prompt in Arize with a name, template, and version metadata. Use this when an agent needs to track a newly written prompt alongside its evaluation history from the first deployment

• list_evaluators: returns the evaluators available in an Arize space, including their type and configuration. Helpful when an agent needs to select the right evaluator before attaching it to a dataset or annotation queue

• list_datasets: retrieves all datasets in an Arize space with their associated metadata. Call this when an agent needs to locate an existing dataset before adding examples or kicking off an evaluation run against it

What makes Merge Agent Handler's Arize MCP server better than alternative Arize MCP servers?

A few factors make Merge Agent Handler the stronger choice for teams connecting agents to Arize:

• Enterprise-grade security and DLP: Merge Agent Handler includes built-in data loss prevention controls that let you block or redact sensitive fields before they reach an agent. For Arize, this means you can prevent proprietary prompt templates, model evaluation results, or internal dataset contents from being exposed to agents that don't require that level of access

• Managed authentication and credentials: Merge stores and refreshes your Arize API credentials on your behalf. You never expose raw API keys to an agent or manually rotate credentials when they change

• Real-time observability and audit trail: Every Arize tool call is logged with the timestamp, tool name, input parameters, and response metadata. You can audit exactly which models, prompts, or datasets an agent accessed without adding any instrumentation to your ML pipeline

• Tool Packs and controlled access: Tool Packs let you define which Arize tools each agent can use. For example, a monitoring agent can access tools like list_monitors and list_models. But it never gets access to create_annotation_queue, dataset write operations, or other tools outside its scope

How can I start using Merge Agent Handler's Arize MCP server?

You can take the following steps:

1. Create or log into your Merge Agent Handler account.

2. Install the Merge CLI by running pipx install merge-api, then run merge configure to link the CLI to your Merge account and merge login to authenticate your session.

3. Register the Agent Handler MCP server with Claude Code by running claude mcp add --transport http agent-handler https://ah-api.merge.dev/mcp, then open Claude Code and run /mcp to confirm agent-handler appears with a connected status.

4. Select agent-handler from the MCP list. This opens a browser window where you select which integrations to authenticate. Choose Arize and complete the API key setup. Merge stores and manages the credentials going forward.

5. Open a Claude Code session and start querying Arize data directly. The first time you use an Arize tool, a Magic Link may appear to complete connector authentication.

If you want to connect Merge Agent Handler's Arize MCP with internal or customer-facing agentic products, you can follow the steps in our docs.