How to use Novu MCP with Claude and Cursor
Your AI client reads your code. Now it can read your notification infrastructure. A 2-minute setup and 20+ tools that turn prompts into real Novu workflow actions.
Your AI client can read your code. Now it can read your notification infra.
Every AI-augmented developer tool has a blind spot: the model sees your code, but not the systems your code talks to.
When you ask Claude to build a notification flow, it generates plausible SDK calls from training data. When you ask it to debug a failed send, it guesses.
That is not AI-native development. That is AI-adjacent.
The Novu MCP server closes the gap for notification infrastructure. Cursor, Codex, Claude Code, and Claude Desktop can now read and write against your live Novu workspace.
Real workflow IDs. Real providers. Real delivery data.
20+ tools, one Bearer token, two-minute setup.
Here is a 45-second walkthrough of the three prompts that make the difference: Watch on YouTube.
What the Model Context Protocol actually changes
MCP is a standardized way for AI clients to talk to external systems.
Instead of every client rolling its own integration for every service, one protocol covers the surface.
A well-designed MCP server is small-grained, typed, and scoped to the intents a model is likely to invoke.
For Novu, that means 20+ tools across seven categories: core operations, subscribers, preferences, workflows, triggering, notifications, and integrations.
Not a REST mirror. Not an SDK dump. Targeted operations a model can pick between without prompt bloat.
The tool surface is a living set - it grows as the product grows.
Build a workflow without leaving your editor
The most obvious use case is also the most time-consuming one in a normal workflow: building a new notification flow.
In a non-MCP world, you open the Novu dashboard, create a workflow, add channel steps one by one, pick providers, configure content, hit publish, then copy the workflow ID back into your code.
A five-minute operation spread across four tabs.
With Novu MCP, you stay in Cursor or Claude Code and type:
Create a workflow called order-shipped that sends an in-app notification when an order ships, then follows up with an email after 24 hours if the user hasn't seen it.
The model calls create_workflow, picks real providers from your environment, wires the delay step, and writes the workflow into your actual Novu workspace.
You get the workflow ID back in the reply.
Five minutes becomes fifteen seconds.
Debug deliveries from chat
The less obvious but more valuable use case: diagnosing failed sends.
Every on-call engineer knows the 2am flow: open the Novu dashboard, filter by subscriber, filter by date, squint at provider responses, cross-reference against the workflow definition.
Novu MCP collapses that into a single prompt:
Why didn't the order-confirmation notification reach subscriber user_789?
The model calls get_notifications and find_subscribers, pulls live delivery data, and answers with the actual cause.
Maybe the email bounced. Maybe the subscriber muted the channel. Maybe the integration is misconfigured.
Whatever the answer, it comes from live state, not a guess.
Bulk operations that used to need a script
A few operations that used to mean writing a small Node script now fit in a prompt:
Bulk trigger the re-engagement workflow for these 500 subscriber IDs.
Cancel the pending order-confirmation event for user_789.
Update subscriber user_456 to mute Slack notifications but keep email and push.
These all hit real MCP tools: bulk_trigger_workflow, cancel_triggered_event, update_subscriber_preferences.
The model picks the tool, passes the right arguments, and reports the outcome.
Install it in two minutes
- Grab a Secret Key from your Novu dashboard.
- Add the Novu MCP server to your client.
For Claude Code, one command:
Other clients
- Cursor: Settings > Tools & Integrations > Add MCP Server
- Codex: edit
~/.codex/config.toml(Codex docs) - Claude Desktop: connect through
npx mcp-remote(Node 18+)
EU environments: use https://mcp.novu.co/?region=eu.
Open-source skills you can fork
We open-sourced the four skills we use ourselves at github.com/novuhq/skills (MIT).
They are the prompts and scoping that turn raw MCP tools into repeatable operations like "debug last night's failed notifications" or "audit my active integrations."
Fork them, remix them, or use them as a reference for skills against other MCP servers.
Missing a tool? Request it on the roadmap
Novu MCP is a living project. The tool set grows with the product, and the next tools come from what the community actually asks for.
If there is a capability you want Novu MCP to support - a tool that is not there yet, an intent the model should be able to express, a workflow your team keeps bumping into - submit it on our public roadmap.
Upvotes and comments directly shape what we ship next.
What is next
Novu MCP is generally available on every plan, cloud and self-hosted. OAuth support is on the roadmap. Rate limits are documented in the platform docs.
Expect the tool surface to keep growing as we see which prompts developers actually write.
If you have been meaning to connect your AI client to real notification infrastructure, this is the version to try.
The model stops guessing and starts reading.
Docs: docs.novu.co/platform/additional-resources/mcp
Skills repo: github.com/novuhq/skills
Roadmap: roadmap.novu.co/roadmap
