Connecting Claude or ChatGPT to Braze With the MCP Server

To connect Claude or ChatGPT to Braze, add Braze's remote Model Context Protocol server as a custom connector and sign in with your own Braze account. The agent then works with your dashboard permissions. Have an admin set up access first, start with reading analytics and content, and treat any tool that changes Braze as a separate, deliberate decision.

Key Takeaways

• Braze's remote Model Context Protocol server connects AI clients such as Claude and ChatGPT to Braze through OAuth, with no API key.
• An agent inherits the connecting person's Braze dashboard permissions, including the right to change data.
• A company admin must turn on OAuth access for the server and grant the server access permission, which is off by default.
• The server's tools don't return data from Braze user profiles.
• Name the workspace in every prompt, and check high-impact answers against the Braze dashboard.
• Agree in writing that a person, not an agent, launches, stops or deletes anything in production.

What is the Braze MCP server?

Braze's remote Model Context Protocol server connects AI agents to Braze analytics and content workflows through OAuth. There's no local package, configuration file or API key to manage, and Braze lists it as generally available in its September 17, 2026 release notes.

An agent can ask about campaign and Canvas performance, explore segments and custom attributes, generate reports, and create content such as email templates and Content Blocks. The tools don't return data from Braze user profiles, and Braze states that tools exposing user-level personal data aren't available.

Braze has verified a list of clients, including Claude and ChatGPT through custom connectors, Cursor, OpenAI Codex, Claude Code and Visual Studio Code. The older locally hosted server from August 2025 is deprecated, and API keys remain in use for the Braze API.

What does an admin need to set up first?

The server won't work until a company admin prepares access. Work through these steps before anyone tries to connect.

1. Turn on MCP OAuth access under Settings, Admin Settings, OAuth.
2. Grant the Use MCP Server permission to each user who needs it, in every relevant workspace.
3. Approve each client's redirect address in the same OAuth settings.
4. Check whether the company uses IP allowlisting, because those customers can't use the server yet.
5. Confirm users have full workspace access, since team-scoped access can stop some tools working.

The server access permission from step two is off by default, which is a sensible starting point. Grant it to a small group first, see how they use it, and widen access once you've agreed how agents should be used.

Every successful sign-in is logged to the security event report. Point your security team at that report before rollout, so they know where to look.

Which endpoint should you use to connect Claude or ChatGPT to Braze?

Braze offers two endpoints, and its documentation states that either one can reach any Braze cluster.

The full addresses to paste into a client are below.

https://mcp.braze.com/mcp
https://mcp.braze.eu/mcp

Choose the endpoint your security team approves, and record that decision. A connection that works isn't the same as a connection your data policy allows.

Why do permissions matter so much?

The agent inherits your dashboard permissions. Braze ties access to each person's Braze account rather than a shared key, so what an agent can see and do mirrors what that person can do in the dashboard.

That includes changes. Braze confirms an agent can change Braze data if the dashboard user has those permissions, and the tool list includes campaign launch and stop tools for customers in the campaign APIs beta.

So the real control is the permission set of each person who connects. Someone with launch rights in production has an agent with launch rights in production, whether or not they meant to give it that.

What should an agent read, and what should it write?

Let agents read freely within their permissions, and write deliberately. Analytics questions, segment exploration and report drafting are the lowest-risk place to start.

Treat writes as a separate decision. Braze's own usage examples include copying email templates from a staging workspace to production, which shows the safer pattern: build and check in staging, then promote.

We recommend a simple written rule for any team that connects. Agents may create drafts and content in staging, but a person launches, stops or deletes anything in production from the dashboard.

How do you keep agent answers trustworthy?

Braze's usage guidance gives four habits, and each one prevents a specific error.

1. Name the workspace in every prompt, especially if you can access several.
2. Give exact date ranges and metrics when you ask for analytics.
3. Ask the agent which tools it used, so you can check how it reached an answer.
4. Compare high-impact recommendations with the source data in the Braze dashboard.

The first habit matters most. An answer about the wrong workspace looks just as confident as one about the right workspace.

How do you revoke access?

Removing a person's dashboard permissions removes those capabilities from any connected agent on the next request. Removing the server access permission ends their agent access completely.

If a token may have leaked or been misused, a company admin can revoke every OAuth token for the company at once. Every connected client loses access and everyone reconnects, so keep that step for a real incident.

What should your security team review?

Data returned by the server flows into the AI client you connect, such as Claude or ChatGPT. The server excludes user-level personal data, but campaign names, content and analytics still leave Braze for that client.

Ask your security team to review each client's data handling before broad rollout, alongside the endpoint choice. That review belongs to your organisation and the client provider, not to Braze.

Which mistakes should you avoid?

Name the mistakes so they're easy to catch. An admin-powered agent is a connection made by someone with broad production rights, so the agent inherits them too.

A wrong-workspace prompt asks a question without naming the workspace and gets an answer about a different one. An unchecked figure is a number an agent produced that nobody compared with the dashboard.

A launch by chat is a production campaign launched or stopped through an agent rather than by a person in the dashboard. Agree in writing that this doesn't happen.

Where does this connect to the rest of the programme?

Agents answer analytics questions best when you know what the numbers mean, as covered in email campaign reporting. Content they create still needs the standards in high-converting email campaigns, and shared blocks are covered in seven ways to deliver dynamic content.

Good answers also depend on clean data, set out in tracking user events in Braze and integrating CRM data into Braze.

As a Braze implementation partner, CustomerIK recommends treating agent access as a permissions project before it's a productivity project. Decide who connects, what they can change, and where writes happen, then connect.

A connection test with an observable result

After the administrator completes setup, add the approved remote endpoint through the client's connector settings and authenticate. Client menus vary, so follow its current instructions. First ask for campaign names and IDs in one named workspace without changing anything.

Compare one result with the dashboard, then request a specific metric and date range. If writes are authorized, test a supported change on a disposable draft and inspect the saved field and edit history. Do not use a production launch as a connection test.

The worked example and test cases are illustrative editorial guidance, not measured client outcomes or a claim that a live integration has been tested.

Frequently Asked Questions

1. Which AI clients work with the Braze MCP server?

Any client that supports remote servers with OAuth can work. Braze has verified Claude and ChatGPT through custom connectors, plus Cursor, OpenAI Codex, Claude Code and Visual Studio Code.

2. Do I need a Braze API key to use the MCP server?

You don't. Sign-in uses your Braze dashboard account through OAuth. API keys still work for the Braze API.

3. Can the MCP server return user profile data?

It can't. Its tools don't return data from Braze user profiles, and tools exposing user-level personal data aren't available.

4. Can an agent change data in Braze?

It can, if your dashboard user has the permissions. Campaign tools such as launch and stop are limited to customers in the campaign APIs beta.

5. How do I remove an agent's access?

Remove the server access permission from the user in the dashboard. If a token may have leaked, a company admin can revoke every OAuth token for the company at once.

Sources

• Braze, The Braze MCP server. Retrieved 18 Oct 2026.
• Braze, Setting up the Braze MCP server. Retrieved 18 Oct 2026.
• Braze, Using the Braze MCP server. Retrieved 18 Oct 2026.
• Braze, Available API functions. Retrieved 18 Oct 2026.
• Braze, What's new in Braze. Retrieved 18 Oct 2026.

CustomerIK is a Braze implementation partner working across onboarding, technical integration, marketing operations and customer data management.

If you're planning to connect AI agents to Braze and want the access rules set first, let's talk.