ChatGPT Connectors

Connect your company’s tools to ChatGPT so people can find answers without assembling a packet of copied documents first. OpenAI now calls these integrations apps, the name that replaces connectors. Its current help pages put app discovery and setup inside Plugins, which bundle apps with workflow instructions. The decision for an admin is which sources, people and actions to allow. OpenAI’s connected-app guide

Start with a managed workspace and a repeated information task. An operations team looking for an approved procedure is a better first rollout than a company-wide assistant with permission to change everything.

This guide reflects OpenAI’s pages checked on October 16, 2026. Menu paths below describe those public instructions. They are not screenshots or claims of access to a private admin console.

Choose Between Live Access, Indexed Search and Research

A live connection looks up information when you ask. An indexed source prepares a searchable collection ahead of time. Think of the difference between sending someone to the records room and giving them a maintained catalog. The catalog still needs to respect who may open each record.

OpenAI has retired individually authorized app sync. Supported indexing is now administrator-managed: an eligible workspace admin connects a source and selects its scope. Content and permission changes can take time to refresh. Connecting your own account for live access does not create that index. Administrator-managed sync

Deep research is a longer investigation across permitted sources. You can review the research plan and check the resulting source links. It uses supported app read actions, not app write actions; an app available in ordinary chat is not automatically available for research. Deep research

Company Knowledge is now a plugin for Business, Enterprise and Edu. Install it, make a supported source available, then invoke it with @Company Knowledge or the tools menu. Installing it does not authorize anyone’s source account. Company Knowledge

The business consequence is a shorter preparation workflow: ask for the brief, inspect its sources, then decide. You still need someone responsible for document quality and access. Moving the retrieval step into ChatGPT does not appoint that owner.

Which Plans Can Use Which Capabilities?

Buy against the capability you need. Access to the plugin directory alone does not establish access to your company’s tools. MCP, Model Context Protocol, is a standard bridge between ChatGPT and tools exposed by an integration server.

The personal-plan rows follow the current plan comparison. Business inclusion follows its plan overview. Enterprise/Edu research access follows the research help page; custom-connection rules follow the MCP help page.

Sync has a separate eligibility test. No personal plan restores individual sync. Google Drive’s current setup page says eligible managed workspaces, without a complete plan matrix. SharePoint explicitly limits administrator-managed sync to eligible Enterprise and Edu workspaces. GitHub has no synced index. Do not infer sync support from a plan’s ordinary app access.

For the four tools in this guide, the narrower documented position is:

There is one further explicit Drive entitlement: OpenAI lists Google Drive in Library for Plus, Pro, Business, Enterprise, Edu and Healthcare on web. That confirms this particular experience, not every Drive action or sync option. Library availability

Decide Who Can See What Before Connecting Accounts

Treat access as several decisions, each with its own owner.

1. Enable the app. Open Admin Console, select the workspace, then Plugins. Inspect the included app; legacy administration is under Manage Legacy Apps. Business controls app availability across the workspace. Enterprise/Edu can assign role access.
2. Review the starting defaults. Business apps are generally enabled. New Enterprise/Edu workspaces start with selected apps enabled; new apps are generally disabled. Existing workspace settings remain in place.
3. Limit capabilities. Where action controls exist, keep only required read capabilities. Review existing actions as well as the policy for future additions. Blocking new actions does not remove old ones.
4. Limit connected identities. Use the app’s domain restriction when available. An empty allowed-domain list permits every domain.

These are separate controls in OpenAI’s admin guide.

Approval prompts are another decision. “Always ask” prompts before reads and changes. “Allow read actions” permits reads automatically and asks before changes. Other settings permit some actions without prompting. These choices govern when ChatGPT asks; they do not give the app more data access. For a reading pilot, I would disable unnecessary writes and choose the read approval policy deliberately. App permissions

When individual authorization is required, each employee connects the provider account that already has the intended access. Some supported sources instead use an administrator-managed connection. Nobody should borrow an administrator’s login to make a test succeed. The connection flow begins in Settings > Plugins, followed by provider authorization where required. Connecting accounts

My suggested acceptance test uses ordinary employees with different source permissions. Give both the same question, including a document one may access and the other may not. Check the linked originals, then repeat after an access change has refreshed. A polished answer is not a permission test.

Three Business Setups You Can Pilot

Each setup starts with a reading workflow. The example prompts are proposed tests, not claims that every object or field is available in every account.

Sales on HubSpot: Prepare an Account Brief

Use this when a seller repeatedly collects account context before calls. HubSpot is the team’s customer relationship management system, or CRM, where customer records live. OpenAI’s Business release notes name HubSpot for chat search and research, but do not supply a complete current field-access contract. Make that uncertainty a setup check, not an assumption.

1. In the managed workspace, find the available HubSpot app or plugin containing it. Review its current description and supported reading capabilities. If absent, stop this pilot at the availability check.
2. Have the admin enable the app and review its access, actions and approval settings using the controls above.
3. Have the seller connect their own authorized account. Review the provider’s requested access before completing sign-in.
4. Try: “Using the HubSpot information available to me, prepare a brief for Acme. Link the records you used, separate recorded facts from suggestions, and identify information you could not retrieve.”
5. Compare the answer with records the seller can open in HubSpot. Ask about a deliberately restricted record and confirm it is not returned before broadening the rollout.

Keep the pilot focused on preparation. Do not promise contact exports, private notes, deal updates or email sending without confirming those capabilities in the current app. The proposed payoff is less context gathering before a call, with the salesperson retaining responsibility for what gets said.

Operations on Google Drive: Answer From Approved Procedures

Use live access for the first narrow test, then consider managed indexing if the workspace qualifies.

1. Enable Google Drive and the required reading actions. Have the Google Workspace admin approve the corresponding authorization scopes, the named permissions Google grants an integration.
2. Connect the operator’s work account through Plugins or Apps, whichever their interface offers.
3. Ask: “Find the approved customer-onboarding procedure in the Operations folder. Cite the source, identify the owner if stated, and flag conflicting versions.”
4. Validate with another employee whose Drive access differs. Check the original documents before using the answer operationally.
5. For eligible managed sync, coordinate a Google service account and domain-wide delegation with your Google admin: an automated identity acts on an administrator’s behalf with approved access. Select My Drive/shared-drive scope, allowed members and account matching. Start with only the approved content needed for the pilot.

Initial indexing can take hours or longer. Follow the provider-specific Drive setup instructions for credentials and permissions rather than improvising a service account.

Google’s consent and ChatGPT’s action settings must agree. Approving one does not resolve a missing permission in the other. Google data controls

For a Microsoft-based operations team, the equivalent starting point is a SharePoint work/school connection. Approval in Entra, Microsoft’s identity-management service, is separate from ChatGPT action enablement and the member’s sign-in. For eligible Enterprise/Edu indexing, choose the approved sites or folders. A narrow indexing scope does not necessarily constrain a separate live connection. SharePoint setup

Engineering on GitHub: Explain an Unfamiliar Code Path

Use this for orientation and investigation in repositories the engineer may already read.

1. Enable GitHub for the intended workspace users and confirm the supported ChatGPT experience.
2. Open Settings > Plugins > GitHub. Complete GitHub installation and authorization, selecting only the pilot repositories.
3. Obtain organization approval if GitHub requires it.
4. Ask: “In the payments repository, locate the refund validation path. Explain the checks and cite the relevant files. Say what you could not verify.”
5. Open the cited code and compare the explanation with the implementation. Test that an unselected repository remains unavailable.

This integration reads repositories on demand. It does not create a GitHub index or push code and pull requests; OpenAI directs code editing and pushing workflows to Codex. GitHub setup and limits

Six Uses Worth Trying, Ranked by Practical Payoff

The strongest first use is the one with a clear source of truth and a repeatable question. These are proposed workflows built on the reading capabilities above, not measured customer results.

For the cross-source handover, verify that both apps and the necessary records are available to the same employee. Do not quietly fill a missing CRM fact with a guess.

Measure the business consequence in a pilot. Record the current sequence, such as find files, read, copy excerpts, write brief, then compare it with ask, verify sources, revise. Track total preparation hours and correction work across the week. There is no defensible universal time-saving number to promise in advance.

If your information already lives in a curated workspace, the tradeoff in Notion AI vs ChatGPT helps decide whether broader retrieval is worth another interface.

Disconnecting Does Not Delete the Work Already Produced

Treat the connection, the conversation and the index as separate records.

These distinctions follow OpenAI’s account-management guidance and sync guidance.

Regular and archived chats remain until deletion or an applicable workspace policy removes them. Deleted chats are scheduled for permanent deletion within 30 days, with documented de-identification, security and legal exceptions. Saved Library files can remain independently of their chats. Chat and file retention

Do not apply that chat-deletion clock to every synced index. Index handling depends on the workspace agreement and applicable controls; OpenAI’s current general sync page gives no universal removal deadline.

Business, Enterprise and Edu app information is not used to train models by default. That is separate from retention. Apps may also receive relevant conversation context, and connected information can inform web searches. Provider privacy terms still apply. Connected-app data use

For personal accounts, model-improvement settings matter. Google-connected data has additional restrictions and exceptions, so avoid a blanket claim that every app’s data follows identical training rules. Google-specific controls

Business administrators do not automatically gain access to members’ ordinary private chats. Compliance Platform access is for eligible Enterprise/Edu workspaces; app log coverage depends on configuration. Admin compliance guidance

Two Things Worth Building Around This

The strongest opportunity is an internal-answer rollout with permission tests. The work a buyer needs is a reliable operating process around their documents.

An operations or IT lead could pay for a small package that identifies approved sources, configures the available connection, creates representative questions and checks answers under different employee permissions. Its smallest sellable version is one department, one source and a repeatable acceptance report. Add an owner for stale documents before adding more integrations.

DataForSEO’s Google estimates retrieved for this guide show 210 monthly US searches for “internal knowledge base software” and 390 for “enterprise search.” Those are category demand signals, not ChatGPT-specific customers. The catch is that a connector cannot repair a badly maintained knowledge base. Your useful work is source ownership and verification; a generic chat interface is easy to copy.

The second opportunity is a sales briefing workflow for a revenue-operations team. Start with an approved HubSpot reading flow, a consistent brief format and a reviewer who can check the source records. Only add a custom integration if the available app cannot retrieve the required information.

The same DataForSEO check returned 480 estimated monthly US searches for “sales automation.” That is broad demand, not proof that buyers want this exact product. The smallest version prepares a reviewable brief for one sales motion. Its catch is incomplete CRM data and changes to app availability; do not sell it as an autonomous seller.

For a custom MCP connection, current Plugins setup uses Add custom MCP server, followed by connection, authentication and review; a Developer Mode toggle is no longer required in that flow. Full workspace support is documented for Business, Enterprise and Edu on web. Custom MCP setup

The Limits That Should Shape Your Rollout

A connected app cannot make unavailable information accessible. Company Knowledge follows the employee’s authorized source access. A useful-looking answer still needs its originals checked. Company Knowledge permissions

Sync is not immediate and is not available everywhere. Personal sync is retired; indexed sources refresh over time. GitHub uses live access. A plan name alone is insufficient evidence for a particular app mode.

Reading does not imply writing. GitHub’s documented app is read-only. Deep research does not use app write actions. Drive and SharePoint can have supported writes, so review their enabled actions instead of assuming every connection is read-only.

An index is not a universal access boundary. Check separate live connections as well as the indexed source. SharePoint does not support personal OneDrive accounts. SharePoint boundaries

Custom integrations need maintenance. Untrusted MCP servers can introduce prompt-injection risks, where retrieved material tries to redirect the assistant. Later tool changes need administrator review; an approved connection is not an automatic approval of every future tool definition. Custom MCP safeguards

My recommendation: prove a reading workflow before authorizing changes. If the answer cannot identify its evidence or the employee cannot verify it, keep the result as a draft.

Does ChatGPT have a knowledge base?

Business, Enterprise and Edu have the Company Knowledge plugin, which uses supported, authorized sources. An eligible managed index is a separate capability. Neither feature means every company file becomes available to every employee. Company Knowledge

How do I create my own knowledge base?

Start with approved documents, named owners and clear source permissions. Then choose an available live connection or eligible managed index. Test real questions and restricted documents before expanding access.

What are some examples of internal knowledge bases?

An operations procedure collection, an engineering runbook library and an approved sales playbook are useful examples. Keep each collection’s owner and audience explicit before connecting it.

On Monday, give the operations lead one approved procedure collection and a colleague with different access. Run the same question, open the sources and record the preparation time. Expand only after the answers and permission boundaries hold up.

If you want this configured around your business tools, we build AI assistants and their integrations.