How to connect Airtable to AI agents with MCP

A practical guide to connecting an Airtable base to an AI agent through MCP: the token, the tools for tables and records, and when to allow writes.

To connect Airtable to an AI agent, run an MCP server that holds an Airtable token and a base ID and exposes operations on tables and records as tools. Start with tools that read, add writes one at a time, and remember that the token decides everything the agent can touch.

What an Airtable MCP server is

MCP is a protocol that lets an AI assistant call functions in another program, called an MCP server. What is MCP? explains it. An Airtable MCP server is the bridge between the assistant and an Airtable base. You ask in plain language, the server calls Airtable, and the answer comes back into the conversation.

Airtable MCP began as a shortcut for myself. During a schema restructuring in an Airtable project I kept going back and forth between the API documentation and my work, copying schema information by hand. I wanted Claude to read that data directly, so I built the bridge. The project is open source under the MIT License.

What it removes is the switch of context. Without it, a question about your data means opening Airtable, finding the right base and table, setting filters or a sort, scanning for the answer and returning to your work. With it you ask from where you already are.

How an AI agent reads an Airtable base through MCP

The server needs two values: an Airtable token and a base ID. You enter them once, at installation. After that the assistant sees the tools that the server declares and decides when to call them during the conversation.

The first version of my server declared these:

list_tables
list_records
get_record
create_record
update_record
delete_record
search_records

The names follow the structure of the data. The base ID says which base the server works with. One tool lists the tables in it. Others list, fetch and search the records of a table. The rest create, update and delete a record.

The server runs locally, on the machine of the person who installed it. How to connect a database securely using MCP covers what changes when a server becomes remote and other people call it. This guide stays with Airtable.

Which Airtable operations to expose as tools

The tool list is the list of things the agent can do, so choose it on purpose. The points below are general advice, not a report of a deployment.

  • Start from the questions people ask. The ones in my own notes are about overdue tasks, open bug reports and stock levels. Tools that read answer all of them.
  • Expose reading first: listing the tables, listing and searching records, fetching one record.
  • Treat every write as its own decision. Creating a record, updating one and deleting one do not carry the same risk.
  • Leave out a tool that nobody has asked for. Every tool you add has to be authorized, tested and reviewed like any other operation.

A longer tool list is not always a better one. What is an MCP gateway? covers what to put in front of a large tool surface.

Airtable tokens and scopes: what the agent can reach

The token is the real boundary. Everything the token can open, the server can open, and the agent can use whatever the server exposes. An instruction in a prompt does not change that.

So decide the access of the token before you install anything. As general advice, create a token for this use only, give it the least access that still answers your questions, and keep it out of prompts and chat messages. If the agent only answers questions, the token should not be able to change anything, as far as Airtable lets you limit it.

This guide does not describe how Airtable limits what a token can do. Read that in the documentation of Airtable and choose the narrowest option that covers the work.

With my server the token is entered once, at installation. The access you grant then is the access the agent keeps.

When to let an AI agent write to Airtable

These are the rules I use.

  • If the base is yours and you accept the risk, a write tool is a fair convenience. In the base of my own project I ask Claude to update the status of a feature to In Progress. The data is mine and so is the risk.
  • If other people depend on the base, start with reading only. Add each write as a separate tool, and apply the checks for writes in the general guide before you turn it on.
  • If the request builds structure, such as tables with linked records, review the plan before it runs. In my notes on the agent, complex migrations still need review.

How to connect Airtable to Claude step by step

  1. Decide what the agent should answer and whether it needs to write at all.
  2. Create the Airtable token with the least access that covers step 1.
  3. Note the ID of the base the agent will work with.
  4. Install the server. In Smithery, search for @rashidazarang/airtable-mcp, click Install and enter the token. The command line form is shown below.
  5. Restart the assistant. The setup notes name Claude Desktop, Cursor, Cline and Zed.
  6. Test through the assistant itself, not only against Airtable, with a question whose answer you already know.
  7. Add write tools last, one at a time.
npx -y @smithery/cli@latest run @rashidazarang/airtable-mcp --config {"airtable_token":"YOUR_TOKEN","base_id":"YOUR_BASE_ID"}

Steps 3 to 5 are the documented installation of my server. The other steps are general advice.

What goes wrong when an agent works with Airtable

  • The token does not survive the trip to the assistant. My hardest bug was authentication. Direct tests against Airtable passed, and the same server failed as soon as Claude called it. The cause was the way the token was handed along the MCP pipeline. Finding it took hours of one evening, and the fix was simple.
  • The agent can execute and cannot reason. Users asked for things such as a project tracker with linked tasks, milestones and assignments. The server failed. It could run each operation, and it did not know the rules of Airtable. Linked records need both tables to exist first. Rollups need specific field types. Lookups cannot reference other lookups.
  • A large change runs without review. Complex migrations need a person to read the plan, and my notes on the agent still list edge cases in formula parsing.

A worked example: from record tools to an agent that knows Airtable

The first server was bare: the tools listed above. For questions it worked well. About two months after I started using it, I was asking Claude for the feature requests with the most user votes, for the number of bug reports still open, and to move a feature to In Progress.

Then I shared it, and people used it in ways I did not expect. Their requests showed the limit. A request to create a project tracker with linked tasks failed, because the server could run commands and could not reason about them.

So I built a second project around the server, an agent for Airtable. It has four parts.

  • An MCP server with a wider set of Airtable operations.
  • A knowledge base: the Airtable API documentation, the formula functions with their edge cases, the field type specifications and rules I learned from years of use. It is split into 56 documentation chunks and indexed for semantic search.
  • An agent that reasons about the request and plans the operations.
  • A context manager that loads the relevant documentation into the context window.

With that, a request for a content calendar produces a plan first: a table of posts, a table of authors, channels, a status workflow in a single-select field, calendar and kanban views, and a formula for the days until publication.

The lesson I take from it: tools let the agent act, and the rules of the product have to come from somewhere else, such as documentation the agent can read.

What this guide shows and what it does not

  • Everything here comes from my own use and from the feedback of people who used an open-source project. I built the server for myself, and the write example runs on the base of my own project.
  • I have no measurements to report: no response times, no accuracy figure and no number of users.
  • This guide does not describe the limits, plans or prices of Airtable. Check them in the documentation of Airtable.
  • A local server with a personal token is not a design for company data, where the boundary has to be enforced outside the model.

Where to go next

Common questions

Can Claude read my Airtable base?

Yes, through an MCP server. I built Airtable MCP, an open-source server that holds an Airtable token and a base ID and lets an assistant such as Claude answer questions about that base in plain language. The token decides what the assistant can reach, so settle that access first.

Do I need to code to connect Airtable to Claude?

Not with the Smithery installation of my server. You search for it, click Install, enter your Airtable token and restart the assistant. A command line option takes the token and the base ID.

Can an AI agent write to Airtable or only read it?

It can do both if the server exposes write tools. The first version of my server could create, update and delete records as well as list and search them. In the base of my own project I ask Claude to update the status of a feature. For data that other people depend on, a write needs validation, authorization and an approval first.

Which AI assistants work with Airtable MCP?

The installation notes for my server list Claude Desktop, Cursor, Cline and Zed. After you install it and restart the assistant, the base is available in the conversation. The protocol does not depend on one model, so the server is not limited to Claude.

Why does an agent fail when it builds a complex Airtable base?

Because running an operation is different from knowing the rules of Airtable. My first server could execute each call and still failed on a request for a project tracker with linked tasks. Linked records need both tables to exist first, rollups need specific field types and lookups cannot reference other lookups. The agent I built later loads the Airtable documentation so that it can plan before it acts.

Work and evidence

MetaMCP as a Cloud Run MCP gateway (opens in a new tab)

My contribution: Built and documented the cloud-facing gateway path, registration artifacts, smoke checks and evidence export.

What this supports: Public implementation account covering transport, caller identity, secrets and tool-call evidence.

What it does not establish: The case study is not an independent security audit or a published concurrency benchmark.

Source reviewed:

Airtable MCP (opens in a new tab)

My contribution: Built an open-source MCP integration for querying and editing Airtable bases.

What this supports: A concrete database-to-agent integration with a linked public repository.

What it does not establish: An integration example, not evidence that arbitrary corporate databases or tenant policies are secured by default.

Source reviewed:

Technical guides