ChatGPT connectors is a useful umbrella term, but it does not identify one specific product or execution model. OpenAI currently describes external-service integrations as apps. Plugins can package apps, skills, or both. Choose a native ChatGPT app when a person wants to search or take a supported action in a conversation, MCP when an MCP-capable client should call tools on demand, and a trigger-based workflow when work must run automatically.
A person checking a project brief during a ChatGPT conversation could use an available app. A one-off request to create a CRM task and post its link could use Zapier MCP. A new lead form that should be classified and routed every time it arrives calls for an automated workflow, with validation before the CRM is changed.
The framework below is an operational comparison, not an official OpenAI or Zapier product classification. Availability and exact capabilities depend on the app, account, plan, workspace, role, region, and enabled actions.
What are ChatGPT connectors now?
OpenAI’s current documentation describes integrations with external services as apps. A plugin can package apps, skills, or both, and the current interface may expose app setup through a Plugins area. “Connector” remains a convenient informal umbrella, not a guarantee that a particular feature, permission, or setup method is available. The official documentation reviewed does not establish a definitive historical sequence in which connectors were renamed and then folded into plugins.
Depending on the app and account, an app may support search, reference, interactive use, or specific write actions. Do not infer that every app can create or update records, or that an app listed for one user is available to every user, plan, workspace, or region. OpenAI’s connected-app guide describes current terminology and account-dependent availability.
Choose by execution mode, not by connector count
Start with how the work should begin. The comparison focuses on the trigger and operational fit rather than trying to compare unlike catalog counts.
| Option | How it starts | Best fit | Key constraint |
|---|---|---|---|
| Native ChatGPT app | A person selects or invokes it in a conversation. | Interactive search, reference, or a supported app action. | Check the specific app, account, and action; availability varies. |
| Custom MCP server | A person invokes approved tools through ChatGPT. | Exposing a proprietary or internal service to an MCP-capable client. | Requires a supported remote-server setup and appropriate workspace controls. |
| Zapier MCP | An MCP client requests an enabled tool action. | One-off or conversational requests that act across connected apps. | The client must meet transport requirements; tool calls use Zapier usage allowance. |
| Zapier workflow | A configured trigger fires. | Repeatable processes that should run without a prompt for each event. | Verify the exact trigger, action, usage rules, and exception handling. |
Zapier documents MCP client examples including ChatGPT, Claude, Gemini, Cursor, and VS Code. Its MCP service requires an MCP client supporting Streamable HTTP and does not support SSE. Zapier currently advertises more than 9,000 apps and 40,000 actions. Those are different, time-sensitive catalog units, and neither count confirms that a particular app exposes the action you need. See Zapier’s MCP overview for current client and tool requirements.
Choose the execution model first. Then verify the exact action, permission, destination, and owner for failures before connecting it.
What connected apps can access, and what approval means
A connected app acts through the authorized account. It does not give ChatGPT access to data that account cannot already access. Read and search capabilities are distinct from write capabilities, so inspect the particular app’s supported actions rather than assuming that an integration can create, send, or update something.
ChatGPT may request approval before reading connected information or completing an action, but confirmation is contextual, not guaranteed before every external change. Particularly risky actions may be blocked instead of offered for approval. Before enabling a consequential write, check the connected identity, requested scope, exact supported action, and whether the approval behavior is appropriate for the risk. Also consider what sensitive data is sent to the provider and retained in task or conversation history. OpenAI explains these conditions in its connected-app documentation.
Set up or remove a ChatGPT app without relying on one menu path
Use the current instructions for the relevant ChatGPT account rather than following a fixed sequence of menu clicks. Open the app or plugin interface available to you, review the listed capabilities, and connect the intended account through its authorization process. Then select or invoke the app in a conversation. Start with a low-risk, read-only request before testing a write.
- Confirm availability. Check that the app appears for your plan, region, workspace, role, and current interface.
- Review capabilities. Identify the exact search, reference, create, or update action needed.
- Authorize deliberately. Use the intended account and review the access requested.
- Test and manage. Verify a read-only result, then test a supported action only if appropriate. Use the account’s current settings or app-management controls to disconnect access.
Workspace approval, account role, plan, and rollout can change both availability and interface labels. Stop if the app, authorization scope, or action differs from the intended use.
When a custom MCP server is the right boundary
Model Context Protocol (MCP) is a protocol that lets a client discover and call tools exposed by a server. It is a protocol layer, not a synonym for a ChatGPT app or a vendor’s catalog listing. A custom server can be appropriate when an organization needs to expose a proprietary service or operation not covered by available apps and is prepared to own the server, its tools, and their access controls.
OpenAI’s custom MCP support connects to remote servers. A local or private server is not connected directly. Where supported, it requires Secure MCP Tunnel. Setup and publishing depend on the account and workspace: use ChatGPT on the web, provide the supported remote-server or tunnel configuration, authenticate, review the exposed tools, and follow workspace publishing and permission controls. The reviewed OpenAI documentation says custom MCP connections are not available on mobile.
Review tool descriptions and scopes before enabling them. Prefer narrow, read-only tools for initial testing, and separate read operations from write or modify operations. Approved tool definitions may be frozen until an administrator refreshes and republishes them, so assign an integration owner to review tool changes as well as initial access. OpenAI’s custom MCP instructions cover connection constraints and safety considerations. For API integrations, OpenAI separately documents remote MCP tools and connector IDs. The availability of a ChatGPT app does not establish that the same integration is available through the API.
Use custom MCP when the service boundary itself needs to be owned. Keep read and write tools separate, publish only the narrow tool set required, and give an administrator or integration owner responsibility for reviewing changes.
Zapier MCP, Zapier automation, and the ChatGPT (OpenAI) app are different
Zapier MCP exposes selected actions from connected Zapier apps as tools an MCP client can request. The client asks for an action, Zapier executes it using the relevant connected app credentials, and the result returns to the client. This is on demand. It does not by itself mean a process will run when a new record arrives.
A Zapier workflow, by contrast, is configured around a trigger and subsequent actions. When the trigger fires, the workflow runs without a ChatGPT prompt for each event. Filters and Paths can apply deterministic conditions around an AI step, so AI can be limited to work such as classifying ambiguous text rather than deciding exact eligibility rules.
The ChatGPT (OpenAI) integration on Zapier is different again. It lets a Zap use OpenAI actions with OpenAI API credentials and prepaid API billing. It does not use a paid ChatGPT subscription login. Do not confuse that integration with Zapier MCP, where an MCP client connects to Zapier-exposed app actions.
Usage depends on the selected product and configuration. Zapier’s current documentation describes task use for MCP calls and different AI by Zapier tier multipliers. Standard, Advanced, Premium, and Bring Your Own Key configurations do not have one universal rate. Check the current task rules and AI tier rules before estimating cost. Zapier’s ChatGPT (OpenAI) setup guide explains the API credential requirement.
Design a safe AI step inside an automatic workflow
Consider this hypothetical lead-routing design, not a vendor-provided template. The system of record is the CRM. A web form supplies a stable source record ID, source event ID, email, consent flag, country, and free-text need. Use ordinary workflow rules for exact requirements, and let AI handle only the bounded interpretation of the free text.
For example, AI can extract the stated need and select exactly one category from an agreed list. The following is a proposed output contract. It is illustrative, not a documented Zapier or CRM feature:
{
"source_system": "web_form",
"source_record_id": "lead-731",
"source_event_id": "submission-2026-11-15-0091",
"classification": "support-operations",
"confidence": 0.82,
"requires_human_review": false,
"prompt_version": "lead-routing-v1"
}Here the row grain is one form-submission event. The combination of source_system and source_event_id identifies that event. It is not a run ID, a CRM contact ID, or an aggregate classification metric. Keep those grains in separate records if the workflow also needs run-level logs or reporting summaries.
For duplicate handling, treat source_event_id as the identity of a form submission, not as a general run ID. Where concurrent runs could process the same event, prefer a destination-enforced unique key or an atomic upsert if the destination supports it. A lookup followed by create can race because two runs may both find no record and then create duplicates. If the destination cannot enforce uniqueness, use a durable idempotency record or a serialized process rather than assuming a generic find-then-create action is safe.
Keep run-level execution logs separate from event-level records. A run ID identifies one workflow execution and can produce several destination actions. If you report classification quality, store the aggregation period, subject, model or provider, and methodology version with the summary rather than treating it as another lead row.
Measure the workflow by successful trigger-to-CRM runs and exception rate, and review classification quality using records checked by a person. Keep original source content separate from generated classification. The workflow owner should be able to find the event ID, validation result, destination record, and failure reason when investigating an exception.
A practical selection and launch check
Before connecting an app or enabling a write, verify the exact operation in the current product documentation. Catalog size alone does not confirm that an app supports the trigger, search, create, or update action your process needs.
- Execution mode: person-led conversation, on-demand MCP tool call, or event-triggered workflow.
- Exact action: confirm the required trigger, search, create, or update is available for the selected app and account.
- Authorization: confirm the connected identity, workspace approval, data sensitivity, and necessary scope.
- Destination identity: validate the record, object, channel, or project identifier before a write.
- Exception ownership: name the person or queue for invalid output, ambiguous matches, permission failures, and partial completion.
- Usage rules: check current plan, task, transport, and AI-tier rules before estimating operating cost.
For interactive use, track successful lookup or action completion. For AI-assisted routing, track classifications that pass review as well as the exception rate. For an automated process, monitor successful trigger-to-destination runs, duplicate attempts, and exceptions.
In short, choose a native app for person-led work in ChatGPT, Zapier MCP for on-demand tool calls from an MCP client, a Zapier workflow for repeatable trigger-led work, and custom MCP when a proprietary service warrants owning its tools and access boundary. If you need help designing and operating a trigger-based process, see ConsultEvo’s Zapier automation services.