The situation
An operations team runs on a handful of Airtable bases. Vendor onboarding, inventory counts, a request intake table, a status board three other teams read. The work is mostly lookup and small edits. Find the record, check a field, update a status, leave a comment. People already have ChatGPT open, so the obvious next step is to connect Airtable to ChatGPT and let it do the lookup.
The part that worries an ops lead is not the reading. It is everything else in Airtable's API surface. Airtable has calls that add a base collaborator, change a share link, start an eDiscovery export and delete records. None of that belongs in a chat window for a team of twelve.
This post sets out how to connect Airtable to ChatGPT through Elaichi. The record tools stay open. Every tool that deletes is blocked with one preset, and the access, export and webhook tools are blocked by name, for everyone except a small admin role. MCP (Model Context Protocol) is the standard way an AI assistant calls tools in other apps.
How do you connect Airtable to ChatGPT without handing over the whole API?
You put Elaichi between them. ChatGPT signs in to one Elaichi address, and Elaichi decides which Airtable tools that person can reach.
Elaichi serves MCP at one endpoint, https://api.elaichi.ai/mcp. The same address serves every organization. Which organization, person and tools a client reaches comes from the OAuth grant behind the token, not from the URL. OAuth here means the sign-in a person approves in a browser, which issues a token to the client instead of a password. The MCP specification sets out that authorization flow, including the discovery document a client reads to find the authorization server (MCP specification, checked November 2026).
The Airtable connector is one Elaichi authors and runs itself. It carries 56 tools and signs in with OAuth, so each person connects their own Airtable account. A call through that connection reaches Airtable as them, with their own base permissions, not as a shared service account.
That matters for the read side. If an ops coordinator cannot open the finance base in Airtable, ChatGPT cannot read it on their behalf. The credential behind the call is theirs. Airtable's own OAuth reference describes the authorization code flow a user approves in the browser (Airtable, checked November 2026).
What does a ChatGPT workspace admin have to do first?
On ChatGPT Business, only admins and owners can create and publish plugins for custom MCP servers for the workspace (checked October 2026, OpenAI). Members then connect from Plugins, subject to workspace permissions.
OpenAI says custom MCP server connections and full MCP support are available for ChatGPT Business and Enterprise/Edu customers on ChatGPT web (checked October 2026, OpenAI). Pro users can connect custom MCP servers with read/fetch permissions.
So the order is simple. An admin publishes Elaichi once for the workspace, then each ops person connects their own account. The admin step is not a per-person step. For the detailed click path, see connecting Elaichi to ChatGPT.
What each person does after that is short. They open Elaichi from their Plugins list and go through Elaichi's own sign-in and consent screen. They pick the organization and approve. Then they connect Airtable inside Elaichi, which opens Airtable's own OAuth window.
Which Airtable tools should the ops team keep?
Keep the record tools. These answer a question or make a small edit, and they are what the team asks for all day.
The reads to leave on:
list_all_airtable_bases, so a person can find which base holds what.list_all_airtable_tables, for the table and field structure inside a base.list_all_airtable_records, the workhorse for "which vendors are still pending".get_single_airtable_record_by_id, for one record in full.list_all_airtable_comments, so the history on a record comes back with it.
The writes to leave on:
create_a_airtable_record, for adding an intake row.update_a_airtable_record_by_id, for a status change or a corrected field.create_a_airtable_comment, for leaving a note on a record instead of editing it.
That set covers lookup, intake and status work. It covers almost nothing else, which is the point. Airtable's own scopes reference lists the permissions an OAuth token can carry, which is a useful second check on what the connection can reach (Airtable, checked November 2026).
Which Airtable tools should be blocked for most roles?
Block every tool that deletes, then block by name the tools that change who can see a base or move data out of Airtable. In Elaichi those are written as block rules on a role.
Start with the preset. When you add Airtable to a block rule and choose Specific tools, Elaichi offers three presets: "Block deleting" (deleting only), "Block all changes" (reading stays on) and "Block everything" (reading too). "Block deleting" catches every Airtable tool Elaichi tiers as a delete in one choice, including the ones nobody thought to list: deleting a record, a view, a webhook, a user, a base collaborator, a workspace collaborator, a share.
Then pick the rest by name. These are not deletes, so the preset does not reach them:
create_a_airtable_base_collaboratorupdate_a_airtable_base_collaborator_by_idcreate_a_airtable_workspace_collaboratorupdate_a_airtable_workspace_collaborator_by_idcreate_a_airtable_interface_collaboratorupdate_a_airtable_share_by_idcreate_a_airtable_ediscovery_exportcreate_a_airtable_webhookcreate_a_airtable_basecreate_a_airtable_table
Three groups sit in that list. Collaborator, invite and share tools change who can see a base. That is the fastest way to turn an internal base into something with a public link. The export and webhook tools move data out of Airtable. The base and table tools reshape the workspace.
Check the connector's own tool list before you save the rule. Airtable has more collaborator, user, invite, webhook and comment tools than any post can name, and a rule only blocks what it names. The preset is what stops the delete half from depending on your memory.
A restriction in Elaichi targets either a role or one member. In Elaichi, restriction targets are role or user only; the organization default is the absence of a rule, which means allow everything. So an ops rollout writes one block rule on the Member role carrying the delete preset plus the named tools, and leaves the admin role without it.
Why a block rule here rather than an allow rule
A block rule leaves everything it does not name reachable. That is right when the dangerous tools are a known list and a preset covers the deletes. An allow rule turns on an allowlist for its whole layer: in Elaichi, once a role carries any allow rule, every connector and tool none of its allow rules names is blocked for that role.
That is a real trap with the Airtable connector. Write an allow rule naming five Airtable read tools and nothing else, and the Member role loses Slack, Notion and everything else at the same time. Those are not on the allowlist either. You would then need allow rules naming every other app the role uses.
An empty allow rule is worse. An allow rule that names nothing blocks every connector and tool for its target. Its summary reads "Allows nothing. Every connector and tool is blocked." The trade-off between the two shapes is covered in the allowlist and blocklist comparison.
For an ops team on Airtable, the block rule is the smaller change. The cost is honest. The preset covers deletes, but a non-delete tool Airtable adds later, or one on this connector that nobody named, stays reachable until someone names it. Review the named half of the block list when the connector changes.
How restrictions behave when a tool is refused
A tool restricted for the caller is left out of the tools Elaichi advertises to an MCP client, and out of a toolbox's runnable tools. The check runs for the person using the toolbox, on the catalog tool's own name and operation.
It is not invisible, though. Elaichi's search_tools searches only the tools the caller may run. It then ranks the caller's restricted tools separately against the same query and returns the matches by name, flagged restricted, with no description, no schema and no way to call them. The person sees that create_a_airtable_ediscovery_export exists and is blocked for them, which beats a confusing silence.
A refused call tells the caller whether the block comes from their role or from their own account, and that an admin can lift it through an access request. The refusal does not name the rule, who wrote it, or what else it covers.
One timing fact to plan around. Role, team-membership, restriction and access-grant changes reach every surface in about two minutes. Write the restriction before you tell the team the connector is live.
Scopes on the ChatGPT side, and what they do not cover
When a person connects ChatGPT to Elaichi, the consent screen draws a checkbox for each scope the client requested. "Use your connected apps" (mcp:tools) is the one that lets ChatGPT run Airtable tools at all.
Here is the part people get wrong. Unticking "Make changes" does not make a connected app read-only. "Use your connected apps" covers running the person's connected tools, writes and deletes included. "Delete things" covers only Elaichi's own deletes. To hold a client to reads, pick toolboxes of read tools at consent, or block the write and delete tools with a restriction.
That is why the Airtable block rule carries the weight here, not the consent checkboxes. If you want a genuinely read-only shape for part of the team, see the read-only rollout pattern.
One more limit, stated plainly. Elaichi's MCP endpoint has no prompt-injection defense of its own. Nothing on the endpoint path scans tool results for injected instructions or second-guesses a call. That defense belongs to the client. Restrictions still hold: a blocked tool stays blocked whatever a tool result tries to talk the model into.
What the audit log shows afterwards
Every connected-tool call that reaches execution writes an audit row. It records the acting person, the toolbox the call went through, the tool's advertised and catalog names, the connector, the connection the call reached, the tool's classification, the surface, the MCP client's id when the surface is mcp, and the outcome.
For a call from ChatGPT, the row carries actor kind user, surface mcp, and the OAuth client's id. The console shows a one-sentence summary, such as the person updating a named record in Airtable through a named connection.
The audit log is not plan-gated. Any organization with an active plan has it, and Elaichi keeps 90 days of history. Members read their own events; audit:view widens that to every event in the organization. For a longer history, forward it to your own log service, which comes with the Black plan.
What the row does not record is every argument. Apart from the one path argument that names the object, a connected-tool row records no argument values, and no argument names or count. So the audit log tells you which record was touched, not the full field payload.
When you do not need any of this
If one ops coordinator wants Airtable in their own ChatGPT and nobody else does, Airtable's own options and a single connection are simpler. The case for Elaichi starts when several people need it. It starts when the block list has to hold for everyone, and when somebody will eventually ask who changed a record last Tuesday.
It also starts when Airtable is not the only app. One address, one set of rules and one audit log is worth something across six apps. Across one, it is overhead.
If the team's real need is a scheduled job that syncs Airtable somewhere else every night, that is a different shape of problem. Automation platform versus MCP access covers which job each one does.
If you are setting up more than Airtable, the first-week governance checklist puts the steps in order, and /use-cases/governance/ shows the wider shape.