ChatGPT Gmail Connector vs MCP

Command references used in this guide: nylas mcp install for local MCP configuration, nylas mcp serve for the STDIO server, nylas mcp status for configuration checks, nylas email search for Gmail retrieval, and the full command reference for exact flags.

What is the difference between a ChatGPT Gmail app and MCP?

A ChatGPT Gmail app is a ChatGPT product connection, while MCP is a protocol for tools and context. OpenAI renamed connectors to apps on December 17, 2025, and its help center lists search, deep research, sync, and write actions as app capabilities.

That naming matters for implementation. A Gmail app in ChatGPT is configured through ChatGPT settings, user OAuth, workspace policy, and the app directory. A custom MCP app is something a developer builds for ChatGPT using OpenAI's app and developer-mode tooling. A local MCP server, such as nylas mcp serve, communicates over STDIO for local clients. Do not assume a local STDIO MCP command can be pasted directly into ChatGPT as a remote app; the transport and hosting model are different. Start with OpenAI's source material: Apps in ChatGPT and ChatGPT developer mode, then compare the protocol layer against the Model Context Protocol specification.

When should you use the ChatGPT Gmail app?

Use the ChatGPT Gmail app when the user wants ChatGPT to reference their Gmail content in a conversation. This is the shortest path for personal search and summarization because it stays inside ChatGPT's app model and does not require a developer to host an MCP server.

The built-in app path is best for 1 user asking questions such as "summarize my latest travel receipts" or "find the thread with the vendor from last week." It is not the best default for a product that must support Gmail, Outlook, Exchange, Yahoo, iCloud, and IMAP with the same workflow. It also puts action availability under ChatGPT plan, region, and workspace settings. OpenAI's help center says admins can configure app availability and action controls, and some app features are plan-limited.

When should you build a custom MCP app for ChatGPT?

Build a custom MCP app for ChatGPT when your product needs approved tools, hosted endpoints, and workspace governance inside ChatGPT. This is a developer path, not a shortcut for running a local command on a laptop.

A custom app makes sense when you own the service behind the tools, can host a remote endpoint, and can document exactly which read or write actions are exposed. OpenAI's app documentation says apps can search third-party services, sync content, and take write actions when configured. For email, that means your tool boundary should be explicit: read message summaries, read full bodies, create drafts, send mail, and search calendar should be separate actions. If you need the same tool surface outside ChatGPT, keep a CLI or API path too.

When should you use local CLI MCP instead?

Use local CLI MCP when the agent runs on the developer's machine and supports STDIO MCP configuration. CLI 3.1.10 supports 5 local assistant targets through nylas mcp install: Claude Desktop, Claude Code, Cursor, Windsurf, and VS Code.

The local path is fast because it writes the client config and starts the Nylas MCP proxy as a child process when the assistant needs it. It is also easy to debug because the same terminal can run nylas email search before asking the agent to search. The command below installs local MCP for Claude Code and verifies that the mailbox search primitive works before handing it to an agent.

nylas mcp install --assistant claude-code
nylas mcp status

# Smoke-test the same Gmail search outside the agent
nylas email search "receipt" --limit 10 --json

See Give AI agents email access via MCP for the local MCP setup details across Claude Code, Cursor, Windsurf, and VS Code.

How do you keep 4 Gmail actions separate?

Keep Gmail agent access split into 4 actions: search, read, draft, and send. A connector that can search does not automatically need to send. A local MCP server that can read one message does not automatically need to download attachments. Separate actions make review and user consent much easier.

This split also maps cleanly to product ownership. ChatGPT's built-in Gmail app is mostly a user-facing context tool. A custom ChatGPT app is a developer-owned tool surface, so its action list should be part of your release notes. Local CLI MCP is a workstation tool, so the assistant configuration and local command pages are the evidence reviewers need. In all 3 cases, send should be treated as a higher-risk action than search.

A good first rollout exposes only search and read. Add draft generation after users trust the summaries. Add send only after the app can show the exact recipients, subject, body, and account that will send the message. That 4-part preview catches most dangerous mistakes before the assistant touches a live Gmail thread.

What should an admin review in 5 minutes?

An admin can review a Gmail connector decision in 5 minutes by checking the account owner, the tool host, the OAuth scopes, the write actions, and the log path. Those 5 facts tell you whether the setup is personal convenience, a workspace integration, or a production tool.

For a ChatGPT app, the admin question is whether the workspace allows the app and whether write actions are enabled. For a custom MCP app, the admin question is who hosts the remote endpoint and who can change the tool definitions. For local CLI MCP, the admin question is whether the user has intentionally configured nylas mcp serve and connected the right Gmail grant. The same checklist works across all 3 paths because it starts from ownership, not branding.

Keep the review result short enough to paste into an issue: "Gmail read-only approved for search and read; send disabled; logs in workspace audit; grant owner is support@example.com." That single line is more useful than a vague approval buried in a chat transcript.

How do you compare read and write actions?

Compare read and write actions by blast radius. Reading 20 message snippets is a lower-risk action than sending 1 email to a customer. ChatGPT apps, custom MCP apps, and local CLI agents should expose those abilities as separate gates.

OpenAI's help center states that apps with external actions must request confirmation before proceeding. Keep that same rule in local agent prompts and scripts: show the recipient, subject, and body before any send command runs.

How do you use CLI commands beside ChatGPT?

Use CLI commands beside ChatGPT when you need repeatable checks that do not depend on chat memory. A developer can run 3 terminal commands to inspect the account, search Gmail, and read one selected message before designing a ChatGPT app or local agent workflow.

The command block below is intentionally read-only. It confirms the active grant, searches for 10 matching messages, and reads a single message by ID. That is the minimum evidence you need before deciding whether to expose search, read, or send as an agent tool.

nylas auth whoami --json
nylas email search "invoice" --limit 10 --json
nylas email read <message-id> --json

Link reviewers to nylas auth whoami, nylas email search, and nylas email read so they can verify syntax without trusting a prose summary.

What should security reviewers check?

Security reviewers should check 6 items before any Gmail agent launches: account ownership, OAuth scopes, write-action confirmation, log retention, revocation path, and provider coverage. A ChatGPT Gmail app answers some of those through ChatGPT app controls; a custom MCP app answers them through app review and server policy.

For local CLI MCP, reviewers should run nylas mcp status, inspect which assistant config was modified, and confirm which grant is active with nylas auth whoami. If the workflow sends mail, require a visible approval step and store only message IDs or subjects in logs. Do not log full message bodies by default, because mailboxes often contain secrets, contracts, invoices, and health data in the same 30-day window.

Next steps

• Give AI agents email access via MCP - configure local MCP for supported assistants
• Google Workspace MCP for AI agents - compare Google's Workspace MCP preview with multi-provider commands
• MCP email server security checklist - review OAuth, local server, and write-action controls
• List Gmail emails from the command line - test Gmail reads before agent setup
• Full command reference - exact syntax for MCP, auth, email, calendar, and webhook commands