RudderStack MCP: Your whole customer data stack, one conversation away

MCP exposes 30+ tools across the customer data lifecycle from pipeline monitoring to transformation authoring and data governance, and we're adding more all the time. Since launching as part of RudderAI in June, we've shipped six point releases, added a web-embedded surface, and connected the server to three new product surfaces. Here's what MCP does today and how teams are already putting it to work.

RudderStack's MCP tools for customer data

When you connect your RudderStack workspace to your AI assistant (Claude, Codex, Cursor) with our MCP server, you unlock full operational superpowers, scoped by your permissions. Here are the five capability groups that are live today:

Here’s what getting started often looks like:

"My Facebook Conversions API destination has failing events. What's causing the errors?"

You’re debugging a pipeline, but instead of manually tracking down the root cause, you enter a prompt, and the agent goes to work. It fetches error samples, identifies the payload shape mismatch, and even suggests a transformation fix, all in one turn. No context-switching between tools, no assembling information from three different screens.

How teams use RudderStack MCP

The entry point is typically debugging. Teams are reducing time to resolution by as much as 90% and catching issues that would otherwise go unnoticed until they show up in downstream dashboards or, worse, surface as customer complaints. Here are a few examples of how customers are using our MCP for debugging and more:

• Catching silent event loss: One team discovered that a consent manager was silently dropping events (the kind of issue that's nearly invisible without actively investigating event flow). Our MCP surfaced it during a routine debugging session, and the team corrected the issue before it reached production dashboards.
• Auditing identity resolution: A DTC brand identified critical overstitching in their profile data through an MCP-powered audit. Profiles that looked correct on the surface were merging users who shouldn't have been merged. It was a silent data quality issue with real downstream impact on personalization and ad spend. They were able to identify and fix quickly thanks to our MCP.
• Operational continuity: One analytics team uses MCP to maintain velocity when team members are out of office. If the person who configured a pipeline is out, our MCP makes their institutional knowledge queryable. The rest of the team can use MCP to get answers about pipeline state and configuration without needing to ask the team member who’s out of office.
• Custom integrations: A customer in the food-tech space built a custom MCP integration for session association via UTMs, extending standard tooling to solve attribution challenges specific to their business.
• Composable AI workflows: An operations platform uses RudderStack MCP as a composable building block inside a broader agentic workflow, enabling their own AI agent to cross-reference data across sources.

Internally: RudderStack MCP is a core component inside Rudder Assist, the internal company AI assistant built by our engineering team. With the MCP, engineers, product managers, and CSMs can simply describe a customer issue in natural language. Then the agent quickly pulls pipeline metrics, inspects destination configurations, and traces event delivery without anyone having to open the web app or troubleshoot in multiple tools. It's connected alongside our warehouse, documentation, and code search tools, so the agent has full context across the stack. That means faster triage, less context-switching, and an 85% merge rate on AI-generated pull requests that flow from Rudder Assist.

An auth model that keeps you, and your data, safe

MCP servers with workspace-level access need authentication that doesn't hand over the keys to the entire org. RudderStack MCP supports two auth methods, each built for a different way of working:

OAuth 2.1: The default for interactive use. When you connect Claude, Cursor, or any MCP-compatible desktop client, OAuth handles the handshake. You authenticate once, the token is scoped to your workspace and permissions, and you're in. This is the path most teams start with.

Bearer token: Built for headless environments. If you're running MCP inside a CI/CD pipeline, an automated agent, a scheduled workflow, or any environment where there's no human to click through an OAuth flow, bearer token auth gives you the same workspace-scoped access without requiring interactive login. Generate a token, pass it in the header, and your automation has full MCP access within its permission boundary. This unlocks MCP as infrastructure. It’s not just an interactive debugging tool, It’s a programmable surface that agents and scripts can call autonomously.

Both paths enforce the same permission model: your agent sees exactly what your account is authorized to see and nothing more. Workspace isolation is enforced server-side regardless of auth method.

There’s more! Agent Skills and the Profiles MCP

The MCP server gives agents tools. Agent Skills give them judgment.

RudderStack Agent Skills is a Claude Code plugin (also available via the Cursor marketplace) that teaches AI assistants the domain knowledge, workflow sequences, preflight checks, and recovery paths for common data engineering work like designing event taxonomies, managing tracking plans, generating type-safe SDKs, and debugging destination delivery failures.

Alongside the core MCP, we open-sourced the Profiles MCP Server, a dedicated server for building identity resolution, feature engineering, and propensity models through conversation. It supports Snowflake, BigQuery, Databricks, and Redshift, and walks users through warehouse discovery, entity definition, and model validation end-to-end.

Getting started

RudderStack MCP works with any MCP-compatible client.Check out the connect guide in our docs, and get started with two simple prompts:

"Give me a health check of my data pipelines over the past week — show error rates, volume trends, and any issues."

"Find event names on my JavaScript source that look like duplicates or typos."

We'll keep shipping point releases on a roughly weekly cadence. Keep up with the latest via the changelog in the repo.