AI-assisted coding has changed how we build software. In ChatGPT-style coder workflows, you describe a feature, watch the code appear, and keep iterating. But the flow almost always breaks the moment you hit infrastructure.
You need a new API key? Break flow.
You need to enable Presence or Message Persistence? Switch tabs.
You want to check whether a user is actually online? Open the admin portal.
That context switch is where momentum dies.
The PubNub MCP Server was built to close that gap. And with the latest iteration now live, it’s evolved from a coding assistant into something much more powerful:
an AI agent that can provision, configure, and operate your real-time backend. It doesn’t just write code anymore. It builds the infrastructure to support it.
From Text Generator to System Operator
If you’ve worked with MCP (Model Context Protocol) before, you know the idea: give your AI assistant tools so it can do real work, not just generate text.
Earlier versions of the PubNub MCP Server were intentionally conservative. They worked, but they were on rails. Apps had fixed names. Keysets came with default configurations. The agent could help, but it couldn’t decide.
That’s changed.
The MCP Server now has deep Admin API control, which means your AI assistant in Cursor, VS Code, Copilot, Claude Code, Claude Desktop, OpenAI Codex, or Gemini CLI can act with real agency using function-style tool calls and guardrails.It can:
Provision infrastructure Create apps and keysets dynamically, named and scoped to your intent.
Configure behavior Enable and tune features like Message Persistence, Presence, Files, and App Context without touching the admin portal.
Debug live systems Query live presence (HereNow and WhereNow) and inspect message history directly from your IDE chat window, even across complex tasks.
Read the manual Pull live SDK and Chat SDK documentation through connectors so your LLMs are grounded in the latest APIs, not AI-generated guesses. This is the shift from “AI that knows the APIs” to “AI that operates the system.” It’s agentic AI in the real world: less demo, more operator.
The Big Unlock: Intent-Driven Infrastructure
Previously, asking a coding agent to “create a new app” often resulted in something rigid: a hardcoded name, a fixed setup, and defaults you’d immediately have to undo.
Now, the MCP Server is dynamic by default.
Because the server exposes granular configuration controls to the AI models, your agent can make decisions based on what you’re trying to build, not just the fact that you asked for an app.
If you say: “Scaffold a HIPAA-compliant healthcare chatbot.”
The agent can infer that it needs to:
Create a new application
Generate a production keyset
Enable the correct security and messaging features
Configure Presence and persistence appropriately
Generate frontend code that matches those decisions
This isn’t a script replaying defaults. It’s an engineer acting on intent, the kind of agentic system startups build whole SaaS offerings around.
Live Docs + Evals = Fewer Hallucinations, Faster Shipping
One of the most frustrating parts of AI development is when a model confidently invents APIs that don’t exist or uses ones that were deprecated years ago. Benchmarks won’t save you if your codebase doesn’t compile.
The PubNub MCP Server tackles this directly. Instead of bundling cached docs, it connects live to the PubNub Docs API. When your agent needs information, it fetches the current source of truth.
So when you ask: “How do I implement typing indicators in the Swift Chat SDK?”
Your AI assistant can:
Pull the latest documentation via connectors
Write code that actually compiles
Provision the backend features required to support it
Help you debug and iterate without leaving the IDE
This closes the loop between documentation, implementation, and infrastructure, which is exactly what multi-agent orchestration is supposed to feel like.
How Developers Are Using This Today
1. One-Shot Chat Prototypes
Spin up a full chat experience with Presence, history, message flow, and live debugging in a single session. No portals. No setup scripts. Just build, verify, refactor.
2. SaaS and OEM onboarding accelerators
If you’re building a multi-tenant product, you already know the pain: provisioning scripts, manual setup, repositories full of edge cases, brittle automation, and pricing surprises as you scale.
With MCP, you can ask:
“Create isolated keysets for these five new tenants and log the credentials.”
Tenant onboarding becomes a prompt, not a project.
3. Operational debugging in production-like environments
Something feels off?
“Check the occupancy of live-stream-01 and show me the last 10 messages.”
You get real-time introspection immediately, without adding dashboards first.
Getting Started
If you’re building real-time features and using an AI-powered IDE, you’re doing more work than you need to if you’re not using this.
Getting started is simple:
Generate an API key in the PubNub Admin Portal
Install the @pubnub/mcp package in your AI tool of choice
Start describing what you want to build, including the setup you’d normally do by hand
While older authentication methods still work, the MCP Server is optimized for API key based authentication. We recommend using it for the best security and experience. Quick links:
PubNub MCP Server docs
https://www.pubnub.com/docs/ai/pubnub-mcp-serverAvailable MCP tools
https://www.pubnub.com/docs/ai/available-mcp-toolsNPM package
https://www.npmjs.com/package/@pubnub/mcp
Your AI already writes your code. Now it can run your real-time backend too.
