BuildersBuildConnect a system

Updated 19 Nov 2026View as Markdown

Connect a system

Lets an agent read from or write to one of your systems, through operations you choose.

Ask your coding agent#

Connect our finance system's HTTP API so the invoice exceptions agent can read invoices and purchase orders. Read only for now.

What gets built#

A connection holds the system's address and credential, with one operation for each action an agent may take. The agent reaches the system only through those operations, and the credential never reaches the model.

TypeUse it forWhat it needs
HTTP APIA system with an HTTP API. Operations come from its OpenAPI description, a documentation page, or your description of each endpointA token, up to three headers, or a key in the address. An API that needs OAuth can't be connected yet, except Google
MCP serverA vendor's remote MCP server. Each tool becomes an operationA token, headers, or an OAuth sign-in. Only remote servers: a local (stdio) server can't be connected
Postgres databaseLookups and joins over more data than a sheet holds. Each query is a fixed, parameterized SQL statementA connection string
Google SheetA small table a person also keeps by hand. Two operations: get_range reads and append_row writesA service-account key, with the sheet shared to the service account's email
Knowledge baseSearching your documents. Mindset doesn't store or index them: they stay on your search serverA remote MCP search server with exactly one tool, rag_search
Google (Calendar, Drive, Gmail)A person's or a team's calendar, files and emailYour own Google OAuth client

Connection types lists every type and its fields.

What you decide#

  • Shared or Individual. One account for everyone, or each person connects their own. See Shared or individual.
  • Which operations exist. Only the ones the agent needs. New MCP tools arrive switched off and offered to nobody: your coding agent validates, enables, offers and grants each one the agent needs, and checks each step.
  • Which are writes. A GET is a read and every other method a write. An MCP tool's marking comes from an admin, then the server's own hints, then a model, and anything still unanswered is a write. A database query's boundary is the login's own privileges, so give a read-only job a login that can't write. Correct any marking that's wrong.
  • Whether a new write needs a person. A read needs no approval. A write is approved once per operation, and with your org's auto-enable setting on (the default) it's enabled with no person involved. See Approve what agents can do.
  • Who can use each operation. An operation is offered to agents only when it and each of its parameters have a description, because the model picks tools from them. A database query starts out available to functions only.

Google runs on your own OAuth client. Before you connect Calendar, Drive or Gmail, create an OAuth client in your own Google Cloud project and set up its consent screen. Your coding agent gives you the redirect URI to add to the client. Google reviews your project, not Mindset:

  • An Internal consent screen (accounts in your own Google Workspace only) needs no verification.
  • An External one needs Google's verification before Calendar or Gmail go to production.

Google's guides: Configure the OAuth consent screen and Create access credentials.

Done when#

  • One real read came back, and you checked a value in it against what the system's owner sees. Enabling an HTTP API read makes that call, and the operation's output shape is taken from the real response. A database query runs inside a transaction that is rolled back. A knowledge base answers a Sample query.
  • You entered the credential on the capture page yourself.
  • Every enabled write is one you meant, and you know whether a person approved it or the org setting enabled it.

See it working#

Live demoHelp widget · LedgerlyA billing help agent reads charges over MCP and stops at a refund nobody has approved. Live demoTrip planner · Wayfarer TravelPlan a Lisbon trip that uses every feature, each one lit on the Features tab as it fires.

In your repo#

Keep each connection and its operations as one file in agents/connections/, beside your agents. The file never holds the credential: a person enters it on a capture page, once per environment. See Keep your agents in your code.

In the Studio#

Connections are created in your AI tool, not in the Studio. Connections → Copy prompt to create opens Build this in your AI tool, with a prompt that asks for a connection in the right environment: paste it into your coding agent and say what to connect. Each connection's page shows whether it's Shared or Individual, and its operations under Operations or Tools on its Overview. When a connected system changes, see Change something that's live.