# Integrations (/docs/integrations)



Other software can work with your workspace in two ways. They look similar from the outside and
they are not interchangeable, because they answer different questions about **who is acting**.

## An AI assistant acting as a person [#an-ai-assistant-acting-as-a-person]

Someone connects their own AI assistant - Claude, or another tool that speaks MCP - to the
workspace. The assistant then acts **as that person**: it can see what they can see and change what
they are allowed to change, and nothing more. A human is in the loop, consenting to the connection
and reading what comes back.

Choose this when the value is a person working faster: asking questions of the data, building an
app by describing it, drafting a Visualisation.

Set it up under [MCP](../integrations/mcp.mdx).

## An unattended system acting on its own [#an-unattended-system-acting-on-its-own]

An ERP, a data warehouse, a nightly job. It signs in with a **credential of its own** rather than
with anybody's account, and it runs on its own schedule with nobody at a keyboard. What it may do
is decided entirely by the access an administrator grants that credential, and every change it
makes is recorded in the audit trail under the credential's name.

Choose this when the value is work happening reliably without anyone present: a recurring import, a
scheduled run, a nightly extract.

Two pages under **API** cover it, one per role:

* **[Credentials](../integrations/api/credentials.mdx)** - for the workspace administrator: issuing
  the credential, choosing its access, rotating its secret, and seeing what it has done.
* **[Overview](./integrations/api/overview)** - for the developer building the integration: minting
  a token, pushing a CSV, running a pipeline and pulling the result. The per-operation **API
  reference** sits beneath it.

## Choosing between them [#choosing-between-them]

| Question                        | AI assistant (MCP)                              | Unattended system (Automation API)        |
| ------------------------------- | ----------------------------------------------- | ----------------------------------------- |
| Who is acting?                  | A person, through their assistant               | A system, under its own credential        |
| Who authorises it?              | The person, by connecting and consenting        | An administrator, by issuing a credential |
| What bounds it?                 | That person's own permissions and AI privileges | The scopes granted to the credential      |
| When does it run?               | While somebody is working                       | On a schedule, unattended                 |
| What does the audit trail show? | The person, and which assistant                 | The integration, by its credential's name |

If both fit, they are not in competition: a workspace commonly has both, for different work.

<Callout type="info" title="The guide and the API reference are English only">
  [Overview](./integrations/api/overview) and the API reference beneath it are not translated. They
  describe a machine contract whose field names, scope names and error codes are English, and a
  translation would eventually describe something the API does not do.
</Callout>