V3.2 answers the other half of the automation question. Nida already runs the business — websites, CRM, voice, email, booking, support. This release is about everything Nida does not own: the tools your team already pays for, and the AI assistants your clients already ask their questions in. The Integrations Hub opens those doors without asking you to paste a credential into a browser tab and hope it stays put.

What V3.2 adds

• A first-class Integrations module with five sections: Discover, My Integrations, MCP Plugins, API & Webhooks, Indie Gems.
• A curated shelf of 17 remote MCP servers — HubSpot, Slack, Gmail, Google Calendar, Stripe, Square, Klaviyo, Notion, Canva, Wix, Webflow, PostHog, Sentry, Neon, Hostinger, plus Zapier and Airtable.
• Two connection styles: a pasted header credential, or a full OAuth 2.1 handshake with dynamic client registration and PKCE.
• Nida as an MCP server at https://www.nida-os.com/api/mcp, so ChatGPT, Codex and Claude can work inside your workspace with scoped tokens.
• Write-only credentials — encrypted on arrival, replaceable or clearable, never readable back.
• One fence for every outbound request Nida makes to a URL a user supplied.

Discover: nothing sits on the shelf that we have not spoken to

Plugin directories rot. An endpoint moves, an auth scheme changes, a vendor ships a beta that half-works, and six months later the marketplace is a graveyard with logos. So the rule for this shelf is blunt: an endpoint holds a row because it answered us. Every entry was probed with Nida's own MCP client — a real initialize followed by a tools/list — before it was written down. Nothing was copied from a directory listing, and nothing is listed with an aspirational badge pretending to be a product.

The catalogue is a floor rather than a ceiling. A workspace can extend it, and the custom path sits in the MCP Plugins section: point Nida at any remote MCP server that answers on the public internet and it will negotiate with that server directly.

Two ways to connect, and why we bothered with the harder one

A token you paste

Zapier and Airtable work exactly the way you would expect: paste the connection token or personal access token once, and the workspace can call their MCP endpoint on its behalf. Zapier is the widest net in the catalogue — one credential, thousands of apps behind it.

OAuth, done properly

For HubSpot, Slack, Gmail, Google Calendar, Stripe, Square, Klaviyo, Notion, Canva, Wix, Webflow, PostHog, Sentry, Neon and Hostinger, Nida runs the real thing: dynamic client registration, an explicit consent screen that names the workspace being connected, PKCE with S256, mandatory rather than optional, exact-match redirect URIs with no wildcard or prefix matching, and a state stored as a hash so a leaked database row cannot be replayed into your account. The flow starts in the browser and finishes on the vendor's own domain — we never see your password there, and we never ask for it.

The credential rule

A credential is write-only. It goes in encrypted, and it can be replaced or cleared — but it never comes back out.

That is not a nicety. It is the difference between a platform you can safely connect things to and one you cannot. A secret that can be read back into a browser is a secret that leaks through a shared screen, an extension, a support session or a log line. So the API never returns it, the UI never renders it, and the status you see is a boolean: connected, or not. Replace the token when it rotates; clear it when the contract ends.

Your workspace, inside ChatGPT, Codex and Claude

The door swings both ways. Nida exposes its own capability surface as an MCP server, complete with OAuth 2.1 authorization-server semantics and a protected-resource metadata document, so an assistant can discover it and negotiate access the standard way. What an assistant gets once it is authorized:

• A workspace snapshot — CRM, websites, automations, bookings, chat, forms and tickets in a single call.
• The live module schema, so it looks up the real columns before writing instead of guessing at field names.
• Module actions across contacts, leads, pipelines and stages, notes, tasks, the media library and websites — including a composite generator that builds a site and all of its pages in one call, surgical page and meta edits, and a template deploy made for agencies rolling one master site onto each client workspace.

All of it runs on scoped tokens against one workspace. Paste https://www.nida-os.com/api/mcp into the connector, choose the workspace, and your assistant is working your actual business instead of a copy of it in a text file.

The fence: we vet where we connect, not how it is spelled

Every outbound request to a URL that a user — or an attacker — supplied passes one gate:

• https only, and no credentials embedded in the URL.
• A hostname that spells a private name is refused outright, with no DNS lookup needed.
• Every resolved address must be public, because a perfectly public-looking hostname can resolve into private space — DNS rebinding, or a name that simply points at a cloud metadata address.
• Redirects are refused, so a fenced host cannot bounce the request to an unfenced one.

The same fence guards OAuth client metadata documents and remote MCP servers, because they are the same class of input: a URL the user handed us. Tenant isolation holds through all of it — plugin state, credentials and flows are workspace-scoped, and a member of one workspace cannot read or act on another's connections.

The small things that make it feel finished

• Variables left the sidebar. It is a read-only reference, so it now lives as a tab inside Settings, where reference data belongs — deep link included, so bookmarks still work.
• API & Webhooks has its own tab and points straight at your keys and event delivery instead of burying them three screens deep.
• The consent-screen dead end is gone. You sign in first, then choose the workspace — the order every other OAuth flow in your life already works in.

FAQ

What is MCP, in one sentence?

MCP (Model Context Protocol) is the open standard that lets an AI assistant call tools a server publishes; Nida speaks it in both directions, so assistants can work your workspace and your workspace can call plugin servers.

Do I have to connect ChatGPT or Claude to use the Integrations Hub?

No. The plugin shelf, credentials and the API & Webhooks tab work on their own. The MCP endpoint is there when you want an assistant working inside your workspace.

Can I connect a server that is not on the shelf?

Yes — MCP Plugins takes a custom remote endpoint. It must be https and publicly resolvable; the fence refuses private, loopback and metadata addresses, and it does not follow redirects.

Where do credentials live?

Encrypted at rest, scoped to the workspace that installed the plugin, never returned by the API. Replace or clear them at any time; nothing else in the workspace can read them.

Does the plugin system work for my clients' workspaces, not just mine?

Yes. Every plugin, credential and OAuth flow is workspace-scoped, which is what makes the Integrations Hub usable on a white-label platform where each client workspace is its own island.

What comes next

V3.1 taught Nida to decide with you; V3.2 lets it work with everything else you already run. The direction from here is depth: more of the catalogue probed and promoted, richer tool surfaces, and the same rule throughout — if we cannot prove it answers, it does not get a row.

Open a workspace and connect your first plugin — or read the V3.1 release notes for the half of the platform that decides what to do next.

← Back to the blog