Skip to main content
API connectors are a Beta feature.
Your database tells the agent what happened. Your product’s API lets it do something about it: flag a transaction, update a record, open a ticket. An API connector turns the endpoints of your REST API into tools the agent can call in chat, and that rules can call as their action. You add it from your API’s OpenAPI spec and approve, endpoint by endpoint, what the agent may call. A change from chat, or from a rule someone sends by hand, waits for that person to confirm it. Scheduled rules and alerts make their changes automatically once they’re switched on, without asking each time. Every change is recorded in API activity.
Project settings, Connectors tab: the MCP servers section above the APIs section, with its Add API button

​
Add an API

API connectors belong to the project, so every workspace in it can use them.
  1. Open Project settings → Connectors and, in the APIs section, click Add API.
  2. Spec: choose From a URL or Upload a file (.json, .yaml or .yml, up to 2 MB), then click Read spec. Upsolve shows the spec’s title, its version and how many endpoints it found. It reads OpenAPI 3.0 and 3.1; a Swagger 2.0 spec is refused with a message saying so.
  3. Who is calling: give the API a Name, check the Base URL, and optionally say When should the agent use this? Then choose how Upsolve calls the API for each user (below).
  4. Endpoints: approve the endpoints the agent may call (below), then Save connection.
For As the signed-in user, a fourth step, Check, confirms your API accepts Upsolve’s token before you save.
The Add an API dialog on its first step, Spec, with From a URL and Upload a file, the OpenAPI spec field and Read spec

​
How Upsolve calls your API for each user

A binding fills a header, query or path parameter from one of the user’s or organization’s properties, the ones you send when you sign a user in. The model can never set it, so each user only ever reaches their own account. Let Upsolve act as your users explains each option and binding in detail.

​
Approve endpoints

Every endpoint starts unapproved, and the agent can only call the ones you approve (up to 200).
  • Each endpoint is marked Read or Write. GET and HEAD start as reads; everything else starts as a write. Change the marking if your API differs.
  • Select all reads approves every read at once. Writes are never approved for you: approve a write only if you’d let the agent and your rules make that change.
  • With bindings, an endpoint that doesn’t take the bound parameter is hidden from end users unless you mark it Global (available to everyone).
  • Endpoints Upsolve can’t call (for example, one whose request body isn’t JSON) say why and can’t be approved.

​
Test the connection

Open the API from the list and use the test panel: pick a read Endpoint to call, who to Test as, and any Arguments. It calls the endpoint once and shows the status and time (for example Connected · 200 in 182 ms). Nothing is saved.

​
Keep it up to date

Click the API in the list to open it. From there:
  • Re-import spec when your API changes. Compare with current shows what was Added, Removed and Changed; Apply takes it. Your approvals and Read/Write markings are kept, and an approved tool is never renamed, so the prompts and rules that use it keep working.
  • Replace key to rotate a server key.
  • Delete to remove the API from the project.

​
How the agent uses it

In a conversation, the agent calls approved reads as it needs them, to look things up in your product. A write is never sent by the agent on its own. The agent prepares the request and shows it as a card, and nothing happens until the person clicks Confirm. Cancel drops it. A request that isn’t confirmed within 30 minutes expires, and if anything about it changed since it was shown, nothing is sent.

​
How rules use it

A rule’s action can be any approved write, and its condition can look things up with approved reads. A scheduled rule or an alert calls the API as the rule’s owner; a button calls it as the person who pressed it, with their own credentials. A write someone sends by hand (a button press, a test or a retry) waits for them to confirm it. A scheduled rule or an alert sends its writes without asking, once its owner switches it on, so approve a write only for rules you’d trust to run on their own. Writes are never retried automatically. See Rules & Actions.

​
Next steps

Let Upsolve act as your users

Each way of authenticating, and how bindings work.

API activity

Every write the agent and your rules sent.