Your integrations
The palette's Your integrations section lists two kinds of ready-made nodes:
- MCP server tools. Every tool of every MCP server registered on the MCP servers page, as
<server> · <tool>. Tools the tool policy blocks for you are left out. - API operations. Every operation of every API added to the project from its OpenAPI 3 spec, as
<API> · <operation>.
Nothing here is a new kind of node. A tool adds an MCP Tool node bound to that server and tool; an operation adds an HTTP Request node with the operation's method and URL. They run, retry, log and replay like any other MCP Tool or HTTP Request node, and a workflow that uses them exports and imports as usual.
Builders and admins see the section. Palette search ranks these entries with the built-in nodes, so typing create ticket finds Ticket Desk · Create a ticket.
Tools of a registered MCP server
- An admin registers the server on the MCP servers page. Its tools are listed when the server is probed.
- Open the palette: the server's tools show under Your integrations.
- Pick one. The MCP Tool node's Arguments form is generated from the tool's input schema: text, numbers, true/false, choices, lists and nested objects get their own fields; anything the form cannot show becomes a JSON field.
Any value may be an expression, such as {{ webhook.ticket_id }}. A number or true/false field takes an expression as is, so it renders as the value, not as text. In a text field, a plain path such as {{ webhook.ticket_id }} is escaped when it renders, so a quote or a line break in the value cannot break the arguments; any other expression is kept as written. The form writes the node's argsTemplate, which is what runs. Picking another tool starts its form empty.
Edit as JSON switches the node back to the plain Tool Args Template field. A node that writes JSON, such as one saved before this form existed, shows Use a form when its tool has a schema; it replaces the JSON (after asking).
Refreshing the tools. The palette reads the tools from the server's last probe. Probes run on the server's schedule; an admin can also refresh one from Add an API… (or the MCP servers page).
Add an API from its OpenAPI spec
- Open the palette and click Add an API….
- Paste the spec's URL, or choose a JSON or YAML file. Optionally give it a name and a base URL.
- Click Add API. Each operation now shows under Your integrations.
Then pick an operation. The HTTP Request node shows the operation (method, path and summary) and a form with its Path, Query and Headers parameters and its Body. The form writes the node's URL, headers and body templates; Request under the form shows what will be sent. A plain path such as {{ webhook.id }} in a path or query parameter is URL-encoded when it renders, and in a header or body text field it is escaped, so a value cannot reach another endpoint or add fields; any other expression is kept as written. Edit the request directly turns the node into a plain HTTP Request node with the same request.
Credentials never live in the spec. Pick a Secret on the node. It is sent the way the spec's security scheme says:
| The spec's scheme | How the secret is sent |
|---|---|
| HTTP bearer, OAuth 2, OpenID Connect | Authorization: Bearer <secret> |
| HTTP basic | Authorization: Basic …; the secret holds user:password |
| API key in a header | in that header |
| API key in the query | as that query parameter (never shown in run logs or error messages) |
| none, or a cookie key | no scheme: a picked secret is sent as a bearer token |
A secret sent in its own header or in the query is never sent to another site: a redirect to another host fails the node instead.
What is supported
- OpenAPI 3.0 and 3.1, JSON or YAML, up to 2 MB and 500 operations. Swagger 2.0 is refused.
GET,POST,PUT,PATCHandDELETEoperations; path, query and header parameters; JSON request bodies (other content types get one text field). Cookie parameters are left out.- Local
$refs (#/components/...). References to other files are treated as any value. - The base URL is the spec's first server. A relative server URL is resolved against the spec's URL. When the spec names no server, set the base URL in the dialog.
Reading a spec from a URL
The URL goes through the same outbound checks as HTTP Request nodes (OUTBOUND_DENY_PRIVATE, OUTBOUND_ALLOWLIST; cloud metadata and link-local addresses are always refused). Redirects are checked again. Reading stops after 15 seconds or 2 MB. A URL with a user name or password is refused.
Refreshing and removing
Refresh reads the spec again from its URL (an uploaded spec asks for the new file) and reports the operations that were removed. A base URL you set stays; one that came from the spec follows it. Nodes already in workflows keep their request and their base URL: when the API's base URL changes, the node's form says so and offers Use the new base URL, so a node's secret never goes to a new host without someone choosing it. A node whose operation was removed still runs, and its inspector says the operation is gone.
A spec URL's query is kept for refreshing but never shown: readers see ?… in its place.
Remove takes the API out of the palette. Nodes made from it keep running.
Who can do what
| Action | Who |
|---|---|
| See Your integrations and add nodes from it | builders and admins with access to the project |
| Add, refresh, rename or remove an API | builders and admins who may edit workflows in the project |
| Register an MCP server, or refresh its tools | admins |
APIs belong to a project: another project's palette does not show them, and they are deleted with their project. The API calls GET /api/catalog, /api/catalog/openapi-specs (see API endpoints).