Connecting Base to your assistant

Base is a database — tables, views, forms, automations — with an MCP server in front of it. Connect it once and your assistant can build a schema, write rows, run SQL against your data, publish a form, and read what comes back.

There is no AI inside Base. The intelligence connects from outside and it is whichever assistant you already pay for. This page covers connecting it, what it can do, and what to do when something goes wrong.

Setup

1. Create an account

Sign in with email or SSO. The free tier is enough to connect an assistant and use every tool — no card, no trial clock.

2. Add the connector

Add Base as a connector in your assistant using this URL:

https://bips.ai/mcp

That endpoint speaks Streamable HTTP. https://bips.ai/mcp/sse is an alias for it, so older configs keep working.

• Claude — Settings → Connectors → Add custom connector, paste the URL.
• ChatGPT — Settings → Connectors → Add, paste the URL.
• Anything else that speaks MCP — same URL. Grok and Gemini both take it.

3. Approve it

Your assistant is redirected to AuthKit, where you sign in and approve the connection. Approving returns you to your assistant and you are done.

There is no API key to create, paste, or leak. Sign-in and connector consent are WorkOS AuthKit (OAuth 2.1, PKCE). Base verifies the token; it does not run the authorization server. You can revoke the connection at any time.

Permissions

A connection carries one of two scopes. Assistants request the read-and-write scope by default; a client that asks for read gets read access only.

Scope	Grants
tables:read	Reading bases, tables, rows, views, comments, history, forms, submissions, automations and delivery logs. Running SQL queries.
tables:write	Everything above, plus creating, updating and deleting. Enforced per operation on the server, not just at the token.

Every tool is tagged so your assistant knows what it is about to do. Read-only tools run without interrupting you. Tools that change or delete data are marked destructive, so your assistant asks first. Deleting a whole base additionally requires an explicit confirmation argument, which the tool description tells the assistant not to set on its own initiative.

What it can do

67 tools. Read and write are always separate tools — there is no catch-all endpoint with a method parameter.

Read-only

list_bases list_tables get_table list_rows get_row get_photo run_query list_views list_comments get_history get_row_history list_forms get_form list_submissions list_automations list_deliveries list_airtable_bases

Creating and sending

create_base create_table create_column create_row duplicate_row create_view add_comment save_version create_form publish_form submit_form release_submission create_automation test_automation retry_delivery import_data migrate_airtable

Changing and deleting

update_table delete_table update_column delete_column update_row delete_row update_view delete_view undo redo update_form delete_form unpublish_form update_automation delete_automation delete_base

Asking questions in SQL

Every table is also exposed as a read-only SQLite view named t_<slug>, with one column per stored field plus id, created_at and updated_at. get_table reports the view name. So instead of paging thousands of rows through your assistant's context to count them:

SELECT status, COUNT(*) FROM t_orders GROUP BY status

Only SELECT and WITH are accepted, only against those views, and the query is rolled back whatever it contains. Formula, lookup, rollup, count and reverse-link columns are computed on read and are not in the views — read those with get_row or list_rows. The formula language is documented at /formulas.md, plain markdown your assistant can read.

Things to ask for

Once connected, in your assistant:

• “Build me a CRM in Base — companies, contacts and deals, linked, with a status field on deals. Then add the six companies from this email.”
• “How many deals are in each stage, and what's the total value of the ones closing this month?” → one run_query call, not forty pages of rows
• “Make a contact form for the Contacts table, publish it, and give me the link.”
• “Anything come in through the signup form this week? Summarise it and flag whatever looks like spam.”
• “Add a webhook that posts to this URL whenever a deal moves to Won, then send a test.”
• “The import doubled every row in Contacts. Undo it.”

The rest of the catalogue, sorted by what you are trying to do, is in the starter prompts.

Troubleshooting

It asks me to sign in again

It should not, day to day. Access tokens last an hour, but your assistant is issued a refresh token alongside them and renews in the background without involving you. You will only see the approval screen again if you go 90 days without using the connection, or if you revoked it.

A tool says forbidden

The connection has read scope only, or your workspace role does not permit that change. Reconnect the connector and check the workspace shown on the approval screen.

A query is refused

The SQL tool reaches only the t_<slug> views. Naming a physical table — or a word that collides with one — is refused by design rather than partly answered. Ask get_table for the view name and query that.

The assistant connected to the wrong workspace

Remove the connector and add it again; the approval screen offers a workspace picker whenever you belong to more than one.

Something else

Email support@bips.ai. Include the tool name and roughly when it happened and we can find it.

Privacy in one paragraph

Base receives the arguments of the tool your assistant calls and nothing else. It does not read your conversation history, your assistant's memory, or files you uploaded to it, and it does not log conversation content. The full detail is in the privacy policy.