Has anyone here started experimenting seriously with the new Braze MCP?

Question

Hi everyone 👋

I’ve been playing with it through Claude Code, and I’m particularly interested in what becomes possible now that the MCP can not only read Braze data, but also modify things like Content Blocks.

I’m currently working on two use cases:

1. Automated campaign reporting

The idea is to have Claude retrieve campaign/Canvas information and performance data from Braze, analyse the results, compare them with previous campaigns or benchmarks, and generate a structured CRM report with key learnings and recommendations.

Ultimately, I’d like those learnings to become reusable context rather than having every campaign analysis live in an isolated spreadsheet or document.

2. Managing A/B tests through Content Blocks

This is the one I find particularly interesting.

Instead of manually rebuilding A/B logic every time, I’m experimenting with centralising reusable experimentation logic inside Braze Content Blocks.

Because the new MCP can modify those Content Blocks, Claude could potentially:

→ Read the existing A/B testing logic

→ Understand the current variants and configuration

→ Modify the Content Block to introduce a new test

→ Update copy / parameters / variants

→ Then analyse the results once the experiment has run

→ Feed the learning back into the next iteration

So you start getting a loop closer to:

build → test → measure → learn → update

with Braze remaining the execution layer and Claude/MCP helping orchestrate the workflow around it.

Still early-stage and I’m trying to understand where this becomes genuinely useful versus unnecessarily complex.

Would love to hear from anyone already using the new MCP:

* What are you using it for?
* Have you started using its write capabilities?
* Has anyone built workflows around Content Blocks specifically?
* And how far are you comfortable letting an AI agent modify Braze configuration?

stefanie.baldwin · Answer

@Axel_Cordonnier

This is a great question- and gets into some technical topics that are a little out of my wheelhouse but I've connected with some folks here at Braze to gather the following details. I'm linking a few other related threads in case they're handy also

First, a correction on the write surface, because it's smaller than the framing suggests. There are 42 tools today. Five of them write: create and update Content Block, create and update email template, and create media library asset. Everything else is read. No campaign or Canvas writes, no user data writes, and no delete tools at all. Content Blocks and templates are the write story right now, which is why your second use case is the one that's actually buildable.

On automated campaign reporting: this works well today. get_campaign_list, get_campaign_details and get_campaign_dataseries plus the Canvas equivalents (get_canvas_details, get_canvas_data_series, get_canvas_data_summary) give you most of what you need, and you can pull get_segment_data_series or get_revenue_series for context. Two things worth knowing: data series come back in daily buckets, so define your comparison windows explicitly or the model will pick something arbitrary, and Canvas data summary and data series answer different questions (totals vs. step level), so be specific about which one you want.

On your reusable context point, the thing that makes this work isn't the MCP, it's where you put the output. In Claude Code, keep a benchmarks file and a learnings file in the project, have the analysis read them first and append to them at the end. That's what turns one-off reports into accumulating context. Without it you get a very nice isolated document every time.

On Content Blocks and A/B logic, one caution: a Content Block update is live the moment it's written. No draft state, no staging, and no delete tool if you make a mess. Everything referencing that block picks up the change immediately. Blast radius is bigger than it feels when you're iterating in a chat window.

The harder problem is measurement. If assignment logic lives inside the Content Block, Braze reporting won't split your metrics by variant, because Braze only reports on variants it created. You'd need to fire a custom event or set a custom attribute per variant and measure off that, and then your analysis loop is reading event data instead of campaign performance. Doable, and some teams do it, but it's the part people underestimate. I'd use native campaign and Canvas variants for the test itself so you keep the reporting, and use Content Blocks for the shared copy and parameters the agent iterates on. Let the agent own the content, not the assignment.

On how far to let an agent modify configuration: the MCP server runs on the permissions of the dashboard user who authorized it. Access is off by default at the company level and gated per user by the "MCP Server" dashboard permission, so the cleanest control is to authorize from an account that has only the permissions you want the agent to have. Beyond that, my rule is that the agent reads, analyzes and drafts, and a human applies anything touching a live asset.

I'd like to hear what's missing for you. If there's a read or a write you keep reaching for and don't have, send it my way, that feedback is useful.

A few other resource in case they're handy:

There's an earlier thread with prompt patterns worth skimming: Braze MCP Server – Prompts and Use Cases,

a Braze Learning walkthrough: Braze MCP Server: Conversational Access to Engagement Data.