Docs / AMS / Connect / Connection types
View as MarkdownConnection types
Work out which of the seven connection types reaches a particular system, what each asks for, and what to do when a system fits none of them.
After this page you can pick the right connection type for each system your agent needs, and know what to do when a system fits none of them.
What this is
When you create a connection you pick what you're connecting to. The type decides what credential it asks for and how its operations are found.
There are seven types. The + New connection picker, titled What are you connecting to?, shows three of them: MCP server, HTTP API and AI model (LLM). To create one of the other four, choose Describe it to the builder and tell the Connection Builder what you need.
| Type | What it reaches | What it asks for |
|---|---|---|
| HTTP API | Anything with an HTTP API | A token, sent as Authorization: Bearer <token> unless you name another header, or up to three headers. The builder can also set up HTTP Basic auth, or no authentication for a public API |
| MCP server | The tools an MCP server exposes | A token or headers, or Sign in in your browser if the server supports it |
| AI model (LLM) | One model, under limits an admin sets | Nothing. Mindset uses its own access to the model |
| Postgres database | A database, through queries you enable one at a time | A connection string |
| Google Sheet | One spreadsheet: read a range, append a row | A Google Cloud service account key |
| Knowledge base | A search server that holds your documents | Whatever that server needs |
| StackOne | HR, recruiting and CRM systems through StackOne | Nothing from you. Each end user signs in through StackOne's Hub |
What each one is for
HTTP API is where most connections end up. The finance system in the invoice workflow is one. The builder finds or writes one operation per method and path, and the reads and the one write become operations on it.
MCP server reaches things Mindset doesn't do itself: reading text from a scanned invoice, producing a spreadsheet, or a product whose vendor ships an MCP server. It's also how you bring across a server someone on your team already runs. Each tool becomes an operation, and from then on it behaves like any other operation: enabled, made available, approved if it writes, and assigned to an agent.
AI model (LLM) is how functions and scripts call a model. An admin picks one model and sets a Size cap (maximum output tokens per call) and PII protection, which is on by default. Every call is audited and its cost logged. A function step or a script phase names a model connection, not a model, so it can't pick a different model than the one it was given.
Postgres database is for lookups and joins over more data than a sheet holds: supplier payment terms and two years of invoice history. Each query is enabled one at a time. Before enabling, Mindset checks your login's privileges and test-runs the query inside a transaction that is rolled back, so enabling it can't change anything. Enabled queries are available to functions unless you also make them available to agents.
Google Sheet is for small tables a person also maintains by hand: who approves supplier queries, or how much variance each supplier is allowed. It has two operations, get_range (a read) and append_row (a write), so it's also a place to drop the month's exceptions for finance to look at.
Knowledge base grounds judgment in your own material. In the invoice workflow, the Assess phase searches the supplier contracts for the clause covering a price difference. A knowledge base connection points at a search server that holds your documents and offers one search tool. Mindset doesn't load or index documents itself, and the Knowledge tab only runs a sample search. Managing sources and tuning search happen on that server.
StackOne reaches HR, recruiting and CRM systems on behalf of each end user. The end user signs in to their own account through StackOne, and StackOne holds that credential, so the outside system sees the end user. Its operations are written for each connection.
How credentials are handled
Whatever the type, Mindset holds the credential and looks it up for each call. The agent never holds it and it never reaches the model.
The host is checked before the credential is looked up. A call aimed at a different host than the connection's is refused before any secret is touched. A call that gets redirected or rewritten on its way out can't send your finance system's key somewhere else.
The outside system sees the connection's credential, not the agent. Mindset's own record names the agent that made the call. With StackOne, the outside system sees the end user's credential.
When a builder needs a credential during a conversation, it opens a form titled Enter a secret value. The model gets back only the name of the field, never the value.
When the system you need isn't listed
Ask three questions, in this order:
- Does it have an API? Most systems do, including many whose vendor sells an integration and never mentions the API underneath. If it has one, use HTTP API.
- Is there an MCP server for it? Vendors and communities publish them for many products. If someone on your team already runs one, bring that across.
- Is it only reachable by logging into a website and clicking through its screens? Then Mindset can't reach it. There is no type that drives a browser. Ask the vendor whether there's an API behind those screens. There usually is.
Limits
- An HTTP API that needs an OAuth sign-in can't be connected yet. Use a token or headers.
- An AI model connection has no cost cap or data residency control. The size cap and PII protection are the limits it enforces.
- A knowledge base can't be filled from Mindset. Its documents live on the search server.
You're done when
- You know which type each system in your workflow uses.
- You have a credential for each one, scoped to what its operations need and pointed at the right account.
- For anything that didn't fit, you know whether it has an API or an MCP server, or that it has neither.