NOXI

Claude connectors, ChatGPT plugins and MCP: what they are, how to add one in Claude, ChatGPT and Gemini, the plan limits and a safety checklist

What the Model Context Protocol is, MCP in eight dates, MCP vs API and plugins, where add-ons live in Claude, ChatGPT and Gemini, setup steps from each help centre, plan limits, connectors by job, a pass/fail safety checklist, troubleshooting and a worked example.

Listen to article 2:26 Try NOXI free
Claude connectors, ChatGPT plugins and MCP: NOXI guide cover about giving your assistant the right keys, with setup steps, plan limits and a safety checklist

Disclosure: NOXI publishes this guide and is one of the tools it mentions; everything about MCP, connectors, plugins and apps applies to any assistant and any tool you connect.

Quick answer: Connectors in Claude, plugins (formerly apps) in ChatGPT and Connected Apps in Gemini let an assistant read from and act in your other tools, and most of them run on MCP, an open standard Anthropic introduced in November 2024. In Claude, add one under Customize > Connectors; the Free plan allows one custom connector.

The names differ by assistant, but the idea is the same: a connector is a key you hand an assistant. The rule this guide is built on: add only connectors from makers you trust, give each one the least access it needs, read every action before you approve it, and disconnect what you stop using. Every plan limit and menu path below comes from the vendors' own help pages, checked on 22 October 2026.

In short

  • MCPAn open-source standard for connecting AI applications to external systems; its own site compares it to a USB-C port for AI applications.
  • ClaudeConnectors from the directory or a custom connector by URL, under Customize > Connectors. Free: one custom connector; Pro, Max, Team and Enterprise: more.
  • ChatGPTPlugins in the sidebar bundle apps and skills; the plugin directory is available across plans. Custom MCP servers: Business and Enterprise/Edu, on the web.
  • GeminiConnected Apps such as Gmail and Calendar; custom apps by MCP server URL need a personal account, age 18+ and the US.
  • SafetyThe MCP specification says tool descriptions "should be considered untrusted, unless obtained from a trusted server", and every vendor tells you to connect only servers you trust.

Key numbers

Nov 2024
Anthropic introduced the Model Context Protocol as an open standard on 25 November 2024 [1]
Dec 2025
MCP was donated to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, on 9 December 2025 [2]
10,000+
Active public MCP servers, as reported by Anthropic at the donation in December 2025 [2]
97M+
Monthly MCP SDK downloads across Python and TypeScript, as reported by Anthropic in December 2025 [2]
1
Custom connectors a Claude Free user can add, per Claude's help centre in October 2026; paid plans can add more [5]
Jul 2026
The current MCP specification revision is dated 28 July 2026 (2026-07-28) [3]

What is MCP? The Model Context Protocol in plain words

MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems, so an assistant such as Claude or ChatGPT can reach data sources, tools and workflows through one common method instead of a custom integration for each pair [4]. Anthropic introduced it on 25 November 2024 [1].

Definition What is the Model Context Protocol (MCP)? MCP is an open-source standard for connecting AI applications to external systems. Using MCP, AI applications can connect to data sources such as files and databases, tools such as search engines and calculators, and workflows such as specialised prompts, so they can access information and perform tasks. Anthropic introduced it in November 2024; since December 2025 it sits in the Agentic AI Foundation under the Linux Foundation.

The official site's analogy is the one to remember: "Think of MCP like a USB-C port for AI applications" [4]. A laptop does not need a different socket for every monitor and charger; an assistant does not need a different integration for every calendar and file store. The tool's maker builds one MCP server, and any assistant that speaks MCP can plug into it.

The specification describes three roles [3]:

  • Host: the AI application you use, such as Claude Desktop, ChatGPT or Gemini.
  • Client: the connector inside the host that keeps one connection to one server.
  • Server: the service that offers context and capabilities, such as a file store, a CRM or a presentation tool.

A server can offer three kinds of things [3]: resources (context and data for you or the model), prompts (templated messages and workflows) and tools (functions the model can run, such as "create a task" or "search my files"). Messages travel in JSON-RPC 2.0, and the design takes inspiration from the Language Server Protocol, which did the same job for programming languages across code editors [3].

What MCP is not: an app store, a product you install or a security guarantee. Consent screens and permissions are the job of the assistant and the server, which is why the safety section below matters more than the setup steps [3].

What is an MCP server, explained simply?

An MCP server is a small program, run by a tool's maker or by you, that tells an assistant what it can do in that tool and carries out the requests. When you add "a custom connector" to Claude or "a custom app" to Gemini, what you paste is the web address of an MCP server [5][15].

Definition What is an MCP server? An MCP server is a service that exposes a tool's data and actions to AI assistants through the Model Context Protocol. It lists what it offers (resources, prompts and tools), receives the assistant's requests and returns results. Remote servers are reached by an HTTPS address; local servers run on your own computer.

A remote server lives at an HTTPS address and usually asks you to sign in; Claude's custom connectors, ChatGPT's custom MCP servers and Gemini's custom apps use these. A local server runs on your own computer; Claude Desktop installs these as desktop extensions [9], and the MCP security guidance warns they "may have direct access to the user's system" [20].

MCP in eight dates: from launch to the Linux Foundation

MCP went from an Anthropic release in November 2024 to support in Claude, ChatGPT, Gemini, Microsoft Copilot and Visual Studio Code within about a year, and in December 2025 it moved to a neutral foundation under the Linux Foundation [2].

DateWhat happenedSource
25 Nov 2024Anthropic releases MCP: the specification and SDKs, local MCP server support in the Claude Desktop apps, and an open-source repository of servers (Google Drive, Slack, GitHub, Git, Postgres, Puppeteer)Anthropic [1]
26 Mar 2025OpenAI's Agents SDK for Python lists "MCP support" as the first key change in release v0.0.7OpenAI on GitHub [14]
1 May 2025Claude "Integrations" with remote MCP servers launch in beta on Max, Team and EnterpriseClaude blog [8]
19 May 2025Microsoft previews MCP support in Windows 11 and publishes its threat list, including tool poisoningWindows blog [21]
3 Jun 2025Claude Integrations expand to the Pro planClaude blog update [8]
14 Jul 2025Claude opens a directory of connectors; remote connectors were then for paid plans onlyClaude blog [9]
9 Dec 2025MCP is donated to the Agentic AI Foundation (Linux Foundation), co-founded by Anthropic, Block and OpenAI; 10,000+ active public servers and 97M+ monthly SDK downloadsAnthropic, Linux Foundation [2][24]
28 Jul 2026Date of the current specification revision (2026-07-28)MCP specification [3]
Timeline of the Model Context Protocol from official sources: Anthropic released MCP on 25 November 2024; OpenAI's Agents SDK added MCP support on 26 March 2025; Claude Integrations launched on 1 May 2025 and reached Pro on 3 June 2025; Claude's connectors directory opened on 14 July 2025; MCP moved to the Agentic AI Foundation under the Linux Foundation on 9 December 2025; the current specification is dated 28 July 2026.
Just over a year from release to a neutral foundation: the plan limits in this guide changed several times along the way, so check the date on any setup page you follow.

One change still confuses searchers: in July 2025 Claude's directory announcement said remote connectors were for paid plans only [9], while the help centre now says web connectors are available to all users and Free users can add one custom connector [5][6].

MCP vs API, and MCP vs plugins: the difference in plain words

An API is one product's own door, with its own rules, that each app must learn separately; MCP is a shared standard, so a tool's maker builds one MCP server and every assistant that speaks MCP can use it. Many MCP servers simply sit in front of a product's existing API and translate.

Plugin, app, connector and add-in are product words; MCP is the protocol underneath many of them.

TermWhat it isWho uses the wordExample
APIA product's own programming interface; each integration is written for that productEvery software companyA developer calls a calendar's API from their own code
MCPAn open standard for how an assistant discovers and calls a tool's data and actionsThe specification, developers, every major assistantOne MCP server works in Claude, ChatGPT, VS Code and Cursor [4]
ConnectorA connection from the assistant to one service, often an MCP serverClaude, Microsoft Copilot, Perplexity, Mistral; Gemini's menuClaude's Google Drive connector
AppA connection from ChatGPT to one service, sometimes with an interactive viewChatGPT; Gemini's "Connected Apps"ChatGPT's Google Drive app [13]
PluginA package of apps and skills for a taskChatGPT; also Claude's unified directoryA ChatGPT plugin that drafts a weekly update from Slack and Linear [11]
Add-inAn extension that runs inside an Office app, not inside the assistantMicrosoft OfficeAn AI add-in in the PowerPoint ribbon

Add-ins live inside Office apps, while connectors live inside the assistant: an add-in brings AI to your open file, and a connector brings your other tools to the AI chat. For the first kind, see our ranking of AI add-ins for PowerPoint, and for Google's editor our guide to AI for Google Slides.

Claude connectors vs MCP: are they the same thing?

A Claude connector is how Claude uses MCP: custom connectors are, in Claude's own help-centre title, "custom connectors using remote MCP" [5]. MCP is the standard; the connector is the saved, signed-in connection inside your Claude account. Some directory connectors are remote MCP servers, and desktop extensions run local servers on your computer [9].

Add-ons in Claude, ChatGPT and Gemini: names, menus and who can add a custom MCP server

Claude, ChatGPT and Gemini all let you connect other tools, but they use different names, put them in different menus and set different rules for adding your own MCP server. The table is a guide, not a score; every cell comes from the vendor's help page opened on 22 October 2026, and Copilot, Perplexity and Mistral follow in their own table below.

AssistantWhat it calls themWhere to find themCustom MCP server: who can add one
ClaudeConnectors (directory and custom); a unified directory also lists skills and plugins [7]Customize > Connectors, or the + button in a chat [6]Free (one custom connector), Pro, Max, Team and Enterprise; on Team and Enterprise an owner adds it first [5]
ChatGPTPlugins, which bundle apps and skills; an app connects ChatGPT to one service [11]Plugins in the sidebar on web, desktop and mobile [11]"Custom MCP server connections and full MCP support are available for ChatGPT Business and Enterprise/Edu customers on ChatGPT web"; Pro users can connect custom servers with read/fetch permissions [12]
GeminiConnected Apps; custom apps by MCP server URL [16][15]Settings > Connected Apps on gemini.google.com (custom apps page); Customize > Connectors (Connected Apps page) [15][16]Personal Google Account, 18 or over, in the US, Keep Activity on; English only; added on the web [15]
Diagram of MCP as a hub: three assistants on the left (Claude with connectors, ChatGPT with plugins and apps, Gemini with Connected Apps) link through one standard, the Model Context Protocol, to MCP servers on the right for files, email and calendar, project trackers and presentations; each server offers resources, prompts and tools.
Three product names, one protocol: a tool's maker builds one MCP server and each assistant reaches it through its own connector, app or custom app menu.

Claude connectors: the directory, custom connectors and plan limits

Claude connectors let Claude "access your apps and services, retrieve your data, and take actions within connected services" [6]. You pick ready-made ones from the Connectors Directory or add any remote MCP server as a custom connector; both are managed under Customize > Connectors.

Definition What is a connector? A connector is a saved connection between an AI assistant and another service, such as a file store, an email account or a project tracker, that lets the assistant read data there and take actions with your permission. In Claude, custom connectors use remote MCP servers.

Four facts from Claude's help pages answer most questions [5][6]:

  • Where they work: directory connectors work on Claude, Cowork, Claude Desktop and Claude Mobile; custom ones are added on the web or in Desktop [5][6].
  • Free plan: custom remote MCP connectors are on Free, Pro, Max, Team and Enterprise, with Free limited to one [6].
  • Organisations: on Team and Enterprise, an owner enables connectors first; each member then signs in [6].
  • Permissions: Claude inherits your permissions from the connected service [6].

Claude connectors list: where the directory is

The Claude connectors list is the Connectors Directory inside Claude: click Customize in the left sidebar, choose the Connectors tab and open Discover [7]. From a chat, select the + button, hover over Connectors and choose Manage connectors [6]. Each connector has a page with its use cases, its read and write capabilities and where it is available [6].

At its July 2025 launch, Anthropic named partner connectors such as Notion, Canva and Stripe, and Figma, Socket and Prisma as desktop extensions [9]. The directory now lists skills, connectors and plugins together, and on Team and Enterprise it also shows items published to your organisation [7].

Claude custom connector on the free plan: one, then upgrade

A Claude Free user can add one custom connector [5]. If you need a second, for example a deck tool and a project tracker, the documented options are to remove the first or move to a paid plan. Claude's pricing page lists Pro at $20 a month on monthly billing, or $17 a month on annual billing, which is $200 billed up front; Max starts from $100 a month (prices exclude tax, checked 22 October 2026) [10].

How to add a custom connector in Claude (how to add MCP to Claude)

To add an MCP server to Claude, go to Customize > Connectors, click "+ Add", choose "Add custom connector", enter a name and the server's remote MCP URL, review the sign-in settings Claude detects, and click "Add" [5]. The steps below follow Claude's help centre as written on 22 October 2026.

On Free, Pro and Max (your own account)

  1. Go to Customize > Connectors.
  2. Click + Add, then Add custom connector.
  3. Name the connector, enter the remote MCP server URL its maker gives you, and continue.
  4. Review the authentication settings Claude detects and change them only if the maker's instructions say so.
  5. Choose how people sign in: Sign in now, Sign in when needed or No sign in.
  6. If you chose a sign-in option, pick the OAuth client option; Claude marks "Use Claude's published identity" as recommended.
  7. Only if the maker asks for one, add a fixed credential such as an API key under Request headers.
  8. Click Add, then sign in to the service when Claude asks.

Claude's help page describes these steps for Pro and Max and gives no separate list for Free, which uses the same Customize > Connectors page within its one-connector limit [5][6].

On Team and Enterprise

An owner goes to Organization settings > Connectors, clicks Add, hovers over Custom, selects Web, completes the same name, URL and sign-in steps and clicks Add. Members then open Customize > Connectors, find the connector (usually labelled "Custom") and click Connect [5]. Owners can set each permission to "Always allow, Needs approval, or Blocked", and members cannot override those limits [6].

How to add MCP to Claude Desktop and mobile

Add custom connectors on the web or in Desktop; they then work on mobile [25]. Remote connections originate from Anthropic's cloud, so the server must be reachable over the public internet; a server on your own laptop or office network will not work as a custom connector [5]. Local servers are a Claude Desktop feature, installed as desktop extensions from the directory [9].

ChatGPT plugins and apps: what they are and how to use them

In ChatGPT, an app connects ChatGPT to a service such as Slack or Google Drive, and a plugin brings apps and skills together for a task; you find both under Plugins in the sidebar, and "The plugin directory is available across ChatGPT plans" [11]. Whether you can install a given plugin depends on your plan, workspace, role, region and the plugin itself [11].

Definition What is a ChatGPT plugin? In ChatGPT's help centre in October 2026, a plugin is a package that can include skills (instructions and workflow guidance), connected apps (links to external accounts, information and actions), app templates for administrators, and extensions such as interactive views. An app connects ChatGPT to one service; a plugin brings apps and skills together for a task.

If you are looking for "ChatGPT connectors" or "ChatGPT apps", they are in the same place. OpenAI's connected-apps page tells you to open Settings > Plugins to connect an app, and notes that "Very old versions of the iOS or Android app may show Apps instead of Plugins" [13]. Its custom MCP page sends administrators to Admin console > Plugins > Manage Legacy Apps for older app settings [12].

How to use apps in ChatGPT

To use an app in ChatGPT, install the plugin that contains it, connect your account if asked, then mention it with @ in a message or choose it from the + menu [11][13]. OpenAI's steps:

  1. Go to Plugins in the sidebar (web, desktop or mobile) and browse the directory [11].
  2. Select a plugin and review its description, skills, apps and setup requirements; a "Desktop only" label means it needs the desktop app [11].
  3. Select Install plugin; if prompted, select Connect and sign in to the provider account that has access to the information you want [11][13].
  4. Review the requested services and permissions and complete the provider's authorisation [13].
  5. In a chat, type @ and pick the plugin, or select + and enter its name. ChatGPT also uses installed plugins on its own when they are relevant [11].

An "OpenAI Verified" badge means OpenAI reviewed the plugin for quality, reliability and usefulness; it "does not replace your organization's privacy, security, or vendor review" [11].

How to add MCP to ChatGPT: custom MCP servers and developer mode

A custom MCP server can be added to ChatGPT by ChatGPT Business and Enterprise/Edu customers on ChatGPT web, from Plugins > + > Add custom MCP server; OpenAI's page says "You don't need to turn on Developer Mode first" [12]. On Business, only admins and owners can create and publish them; on Enterprise/Edu, admins can grant the permission to selected members [12].

OpenAI's steps, as written on 22 October 2026 [12]:

  1. Go to Plugins.
  2. Select the plus button, then Add custom MCP server.
  3. Enter a name and, optionally, a description.
  4. Under Connection, enter the server's public HTTPS MCP endpoint URL (or choose a Secure MCP Tunnel).
  5. Configure authentication for the server.
  6. Review the risk warning and select I understand and want to continue.
  7. Select Create as a plugin and review the discovered tools and metadata.

The same page settles three common questions [12]: custom MCP servers are web-only; ChatGPT connects to remote servers, not directly to a local one; and "Pro users can connect custom MCP servers with read/fetch permissions", while full MCP is for Business and Enterprise/Edu. Tutorials that start with "turn on Developer mode" describe an older path.

Gemini Connected Apps and custom apps (MCP): US, 18+, personal accounts

Gemini connects to other apps through Connected Apps, such as summarising Gmail or creating calendar events, and it also accepts custom apps by MCP server URL if you are 18 or over, in the US, signed in with a personal Google Account and have Keep Activity on [15][16].

People still search for "Gemini extensions"; Google's help pages today say Connected Apps. With Keep Activity off, Connected Apps are unavailable on gemini.google.com, iOS and smart watches, and on Android only Device assistance, Phone, Messages and WhatsApp remain; type @ in the prompt box to pick an app for one request [16].

How to add a custom app (MCP server) to Gemini

Google's steps for a custom app, which you can only add in the Gemini web app [15]:

  1. On a computer, go to gemini.google.com.
  2. At the bottom, click Settings > Connected Apps. If you don't see it, click Personal Intelligence > Connected Apps first.
  3. Under Custom apps, use the "Add a custom app link to get started" field, or click Add a custom app.
  4. Enter the app's MCP server URL. If the server doesn't support Dynamic Client Registration, open Advanced features and enter your credentials.
  5. Click Next and follow the on-screen instructions.

For now, custom apps are English-only and available only in the Gemini mobile app and the web app; once connected on the web, they work in both [15]. Google is direct about the risk: it "does not control, monitor, or secure third-party MCP servers", and custom apps "may request data beyond what is strictly needed for your request" [15].

Copilot, Perplexity and Mistral: connectors in one table

Microsoft Copilot's consumer app connects to a fixed list of services, while Perplexity and Mistral let you add your own remote MCP server as a custom connector, with admin controls for organisations. The table summarises each vendor's help page; plan names appear only where the page names them.

AssistantWhat you can connectHow to add oneNote from the vendor
Microsoft CopilotOneDrive, Outlook.com (email, calendar, contacts), Google Drive, Gmail, Calendar and Contacts, on Microsoft 365 Personal, Family, Premium or Pro [17]Open the connectors list from your Profile, select Connect and authorise the account [17]Copilot "only works with content your account already has permission to view"; for Microsoft services the account must match the one you sign in with [17]
PerplexityCustom remote connectors to any MCP server over HTTPS [18]Plugins in the sidebar, add a custom connector, choose Remote, enter a name and the MCP server URL, accept the risk acknowledgement, Add [18]"Perplexity cannot verify third-party MCP servers"; "Only use connectors from sources you trust" [18]
Mistral VibeDirectory connectors or custom MCP connectors [19]Connectors > + Add Connector > custom option; enter a name and the server URL, then Connect [19]"Keep write functions on manual approval until you're confident in the Connector's behavior" [19]

The best Claude connectors for everyday work, by job

The best Claude connectors remove a copy-and-paste step you do every week: files, email and calendar, notes, design and presentations. Start with one connector per job, from a tool you already use.

The table names only examples the vendors list on their own pages; the directory in your own account is the current list.

JobWhat a connector doesExamples named on official pagesAsk it
FilesFinds and summarises documents you can already openGoogle Drive (Anthropic's first servers [1]; ChatGPT [13]; Copilot [17]), OneDrive (Copilot [17]), SharePoint (ChatGPT [13])"Find last quarter's budget file and list the three biggest changes."
Email and calendarReads threads, drafts replies, creates eventsGmail and Calendar (Gemini [16]; Copilot [17]), Outlook (ChatGPT [13]; Copilot [17])"Summarise unread emails from the client and propose two meeting times."
Team chatSearches channels and messagesSlack (Anthropic's first servers [1]; ChatGPT [11]), Microsoft Teams (ChatGPT [13])"What did the design channel decide about the launch date?"
Notes and projectsReads pages and tasks, files new onesNotion (Claude directory launch [9]; MCP site [4]), Linear (ChatGPT example [11])"Turn these meeting notes into five tasks with owners."
DesignOpens a design or creates a draftCanva and Figma (Claude directory launch [9])"Pull the colours from our brand file into a one-page summary."
CodeReads repositories and pull requestsGitHub (Anthropic's first servers [1]; ChatGPT event-triggered tasks [13])"Summarise the open pull requests and what each changes."
PresentationsBuilds a designed deck from the chat and returns the filesPresentation tools that publish an MCP server, such as NOXI (see Build it below)"Turn this research summary into a 10-slide deck."

What are the best MCP servers for everyday users?

For everyday users, the best MCP servers are the official ones from the maker of a tool you already use, found in your assistant's own directory rather than a list of unknown servers. OWASP's advice applies to everyone: "Only install MCP servers from trusted, verified sources" [23].

Are MCP connectors safe? What a connector can read, and the checklist

MCP connectors are as safe as the server you connect and the permissions you grant: the protocol itself "cannot enforce these security principles", so trust in the maker, the least access and your approval of each action do the work [3]. Every vendor in this guide says the same thing in its own words: connect only servers you trust [5][12][15][18].

What a connector can read and do

A connector can read what your account in the other service can read, and do what you allow it to do there. Claude's help says connectors only sync content you have permission to view, and Microsoft says Copilot "only works with content your account already has permission to view" [6][17]. Two consequences follow:

  • Your access is the ceiling, not the default. Connecting your work drive with full access gives the assistant everything you can open, including folders you forgot about. Where a sign-in screen offers narrower scopes, choose them; Claude's help says to "deny access if requested permissions seem unnecessary" [5].
  • Data leaves the assistant. Claude notes that connected services process data on their own infrastructure under their own terms [6]; Google says that once data reaches a custom app, it is "subject to that app's privacy practices" [15]. Read the connector maker's privacy policy as you would any other vendor's; for what the assistant itself keeps, see whether ChatGPT, Claude and Gemini train on your chats.

The four risks, in plain words

  • Tool poisoning: "Malicious instructions hidden in tool descriptions, parameter schemas, or return values that manipulate the LLM's behavior" (OWASP) [23]. This is why the specification says tool descriptions "should be considered untrusted, unless obtained from a trusted server" [3], and why Claude warns that "Malicious MCP servers may include hidden instructions" [5].
  • Prompt injection from content: a document, email or web page the assistant reads contains instructions aimed at the assistant. Microsoft calls it cross-prompt injection: content that "can override agent instructions" [21]. OpenAI lists prompt injection as a risk of untrusted servers [12].
  • Rug pulls: "A server changes its tool definitions after initial user approval" (OWASP) [23]. Claude's help tells you to monitor tool behaviour changes [5]; ChatGPT workspaces use a "frozen" snapshot of a server's tools until an admin publishes an update [12].
  • Confused deputy: a flaw in a server that connects you to a third service, which can let an attacker obtain access codes without your consent [20]. The server maker fixes it; your defence is a trusted maker.

For security teams, OWASP keeps an MCP Top 10, currently a beta, with risks such as privilege escalation via scope creep and shadow MCP servers [22].

Connector safety checklist (pass/fail)

  1. Trusted maker.Pass: the server address comes from the tool's own website or your assistant's directory. Fail: a URL from a forum post or a stranger's message.
  2. Least access.Pass: you granted only the scopes the job needs and denied the rest. Fail: "full access" accepted to save a click.
  3. Read before you approve.Pass: you read every write action (send, create, delete, share) before approving it. Fail: approving a batch of actions unread.
  4. Unneeded tools off.Pass: tools you don't use are turned off (Claude: the "Search and tools" menu [5]). Fail: every tool of every connector on in every chat.
  5. Sensitive data stays out.Pass: your company policy allows this data to go to this service. Fail: client files sent to a server nobody reviewed.
  6. Review every quarter.Pass: you disconnect connectors you have not used in 90 days. Fail: a connector from a past project still signed in.
Safety checklist for AI connectors in four cards: add only connectors from makers you trust; give each one the least access it needs; read every action before you approve it; disconnect what you stop using. A footnote quotes the MCP specification: tool descriptions should be considered untrusted unless obtained from a trusted server.
Four habits cover most of the official warnings: trust the maker, limit access, read before approving and disconnect what you stop using.

For the wider question of which documents belong in any AI tool, our template for confidential material in AI presentation tools walks through a classification and a sign-off step.

Troubleshooting: Claude connectors not working, ChatGPT apps not showing

When a connector is missing or fails, the cause is usually the plan, a workspace policy, the region or account type, an activity setting, or a server the assistant cannot reach.

SymptomLikely causeWhat to check
Claude: can't add a second custom connectorFree plan limit of one [5]Remove the first, or move to a paid plan
Claude: connector missing at workTeam or Enterprise owner has not enabled it [6]Ask the owner to add it under Organization settings > Connectors
Claude: custom connector won't connectServer not reachable from Anthropic's cloud [5]The URL must be a public HTTPS address, not a local or office-network server
ChatGPT: "Disabled by admin"Workspace policy [13]Ask an administrator to review the app or plugin in Admin console > Plugins
ChatGPT: no "Add custom MCP server"Plan or role [12]Business or Enterprise/Edu on the web, with the "Create plugins with MCPs" permission
ChatGPT: app shows as "Apps"Very old mobile app version [13]Update the iOS or Android app; the menu is now Plugins
Gemini: no Custom apps sectionAge, country, account type or language [15]18+, in the US, personal account, English, on gemini.google.com
Gemini: Connected Apps goneKeep Activity is off [16]Turn Keep Activity on; custom apps need it too [15]
Copilot: can't find my filesAccount mismatch [17]The Microsoft account in Copilot must match the one for OneDrive or Outlook

If none apply, disconnect, reconnect and sign in again, then contact the server's maker.

Worked example: a three-person studio connects two tools to Claude

Example Studio Kanata is a fictional three-person design studio. It wants Claude to reach two tools through custom connectors: its presentation tool, NOXI, for client decks, and its project tracker, which publishes its own remote MCP server.

Step 1: the plan limit. Claude's Free plan allows one custom connector [5]. The studio needs two, and 2 > 1, so each of the three people needs a paid plan to keep both connectors.

Step 2: the cost. Claude Pro on annual billing is $200 a year per person, shown on the pricing page as $17 a month; on monthly billing it is $20 a month [10].

BillingPer personStudio of 3, per yearWorking
Annual$200 a year$6003 × $200 = $600
Monthly$20 a month$7203 × $20 × 12 = $720
Difference$40 a year$120$720 − $600 = $120; 120 ÷ 720 = 16.7%

Annual billing saves $120 a year, 16.7% of the monthly total, before tax [10].

Step 3: sign-ins and keys. The deck tool's connect page states "1 sign-in per assistant" and "0 API keys or passwords to paste" [25]. For one assistant and three people that is 3 × 1 = 3 sign-ins and 0 keys. The tracker uses its own sign-in, so the studio counts 3 more: 6 sign-ins in total, and nothing secret pasted into a chat.

Step 4: access. The deck connector's consent screen lists reading presentations and creating, editing and exporting them [25]. For the tracker, the studio grants task access only and turns off its delete tool in Claude's "Search and tools" menu [5].

Step 5: upkeep. A connection to the deck tool "stays active for up to a year while you use it at least once a month" [25]. The weekly deck review uses it 52 times a year, above the 12 monthly uses needed, so each person signs in again at most once a year. They connect on 1 December 2026 and book a 90-day access review for 1 March 2027.

Here are two slides from this article's worked example, as NOXI built them, unedited.

NOXI slide: what each assistant calls its add-ons and where to find them, Claude connectors under Customize then Connectors, ChatGPT plugins in the sidebar, Gemini connected apps and custom apps in Settings, most of them running on MCP.
The names slide: one table for three assistants' add-on layers.
NOXI slide: four rules for every connector, trusted maker, least access, read then approve, and disconnect any connector not used in 90 days.
The rules slide: four checks before and after you connect anything.

3 prompts for building a deck from an assistant with NOXI connected

Once a deck tool is connected, name it in the request so the assistant picks the right connector, and give the slide count, the audience and the facts to use.

Prompt 1 · A training deck from this guide

Create a 9-slide NOXI presentation for [our team of 12] titled "Connecting our AI assistant safely". Slides: 1 title and date [25 November 2026]; 2 what MCP is, with the official definition and the USB-C analogy; 3 where add-ons live in Claude, ChatGPT and Gemini; 4 our plan limits [Claude Free allows one custom connector]; 5 the connectors we use [list]; 6 the four safety rules; 7 the four risks in plain words; 8 our cost [numbers]; 9 next steps with dates. One idea per slide, every number with its unit.

Prompt 2 · A client deck from your notes

Turn the notes below into a 10-slide NOXI presentation for [a client board meeting on 3 December 2026]. Keep every figure exactly as written, put one finding in each slide title, add a table for [the budget comparison], and end with a slide of three decisions we need. Notes: [paste].

Prompt 3 · A follow-up edit, then the files

In my latest NOXI deck, turn slide 4 into a three-row table, shorten every slide title to under 10 words, and make slide 9 a timeline of [our three dates]. Then export the deck as PowerPoint and PDF.

For more prompt patterns, see our 40 prompts for presentations.

Build it: connect NOXI to Claude and ask for a deck

NOXI publishes one MCP server address for every assistant on its connect page: https://mcp.noxi.ai/mcp. You sign in to NOXI in your browser, so there is no API key to copy [25]. The steps below are the Claude path, using Claude's help-centre wording for the menu; noxi.ai/connect also lists ChatGPT, Cursor, Claude Code and Codex as MCP clients.

Step 1: Copy the server address from noxi.ai/connect

Open noxi.ai/connect and copy the MCP server address, https://mcp.noxi.ai/mcp. Every assistant connects to the same address.

Step 2: Open the custom connector form in Claude

In Claude, go to Customize > Connectors, click + Add, then Add custom connector. The connect page calls the same place Settings → Connectors.

Step 3: Name the connector and paste the address

Enter NOXI as the connector name and paste https://mcp.noxi.ai/mcp as the remote MCP server URL. Keep the sign-in settings Claude detects: the connect page says no client ID is needed and there is no API key to paste.

Step 4: Sign in to NOXI and approve access

Select Add, then Connect, and sign in to your NOXI account in the window that opens. The consent screen lists what the connector may do, such as reading your presentations and creating, editing and exporting them; approve it.

Step 5: Ask for the deck in the chat

With the connector on, name NOXI in your request, for example "Create a 10-slide NOXI presentation about our Q3 product launch", or paste Prompt 1. When NOXI creates or changes a deck, the PDF and PowerPoint files arrive in the chat.

Step 6: Open the deck in NOXI to edit, present or export

Open the deck in NOXI to edit any slide or the whole deck by chat, present it full-screen, share a view-only link, or export an editable PowerPoint (.pptx) with real text boxes and embedded fonts, or a PDF. You can disconnect the connector at any time in Claude's connector settings.

The connection works only with presentations in the NOXI account you connected [25]. For decks made in the assistants themselves, see our comparisons of Claude for presentations and ChatGPT for presentations, and for polishing an existing file, our scored list of AI tools that make a PowerPoint look better.

Frequently asked questions

What is MCP in AI?
MCP, the Model Context Protocol, is an open-source standard for connecting AI applications to external systems such as files, databases, search tools and workflows. Anthropic introduced it in November 2024, and it now sits in the Agentic AI Foundation under the Linux Foundation. Its official site compares it to a USB-C port for AI applications.
What are Claude connectors?
Claude connectors let Claude access your apps and services, retrieve your data and take actions in them with your permission. You add ready-made ones from the Connectors Directory or any remote MCP server as a custom connector, both under Customize > Connectors, and they work on the web, Claude Desktop and the mobile apps.
Can I add connectors on Claude's free plan?
Yes. Claude's help centre says web connectors are available to all users, and custom remote MCP connectors are available on Free, Pro, Max, Team and Enterprise, with Free users limited to one custom connector. If you need a second custom connector, remove the first or move to a paid plan such as Pro.
How do I add a custom connector to Claude?
Go to Customize > Connectors, click + Add and choose Add custom connector. Enter a name and the remote MCP server URL, review the sign-in settings Claude detects, choose how people sign in, and click Add. On Team and Enterprise, an owner first adds it under Organization settings > Connectors, then members click Connect.
What happened to ChatGPT apps and connectors?
They now sit under Plugins. In OpenAI's help centre, an app connects ChatGPT to a service such as Google Drive or Slack, and a plugin bundles apps and skills for a task. You find them under Plugins in the sidebar, and very old versions of the mobile apps may still show Apps instead.
How do I add an MCP server to ChatGPT?
On ChatGPT Business or Enterprise/Edu, on the web, go to Plugins, select the plus button and choose Add custom MCP server. Enter a name and the server's HTTPS endpoint, configure authentication, accept the risk warning and select Create as a plugin. OpenAI says you don't need to turn on Developer Mode first.
Does Gemini support MCP?
Yes. Gemini accepts custom apps by MCP server URL, added at gemini.google.com under Settings > Connected Apps > Custom apps. You must be 18 or over, in the US, signed in with a personal Google Account and have Keep Activity on. For now, custom apps are English-only, in the Gemini web and mobile apps.
What is the difference between MCP and an API?
An API is one product's own interface, so every app must build a separate integration for it. MCP is a shared standard: a tool's maker builds one MCP server, and any assistant that speaks MCP, such as Claude, ChatGPT or Gemini, can use it. Many MCP servers sit in front of an existing API and translate requests.
Are MCP connectors safe?
They are as safe as the server you connect and the access you grant. The MCP specification says tool descriptions should be treated as untrusted unless they come from a trusted server, and every major assistant tells you to connect only servers you trust. Grant the least access, read every action before approving it, and disconnect unused connectors.
Can I make a presentation from inside Claude?
Yes. Connect a presentation tool that publishes an MCP server as a custom connector, then ask for a deck in the chat. For example, connect NOXI with the address on noxi.ai/connect, ask for a 10-slide deck, receive the PDF and PowerPoint files in the chat, and open the deck in NOXI to edit or present it.

Sources & methodology

How this guide was made: every plan limit, menu path and quotation comes from the vendor's own help centre, pricing page, blog, release notes or specification; OpenAI's help pages and noxi.ai/connect were read in a browser. The checklist and the table by job are editorial recommendations. Studio Kanata is fictional; every calculation was checked twice. All sources were checked on 22 October 2026.

Connect NOXI to your assistant with the address on noxi.ai/connect, ask it for a deck, then open the deck in NOXI to edit it by chat, present it full-screen or download an editable PowerPoint or PDF.

Written by Aidar Akmaev

Founder and designer of NOXI, an AI presentation maker for professional, consulting-grade decks. NOXI is one of the tools mentioned here; see the disclosure at the top.