Create Accountor Login
Docs

Connection Actions

Part of the MountDev Ops Integrations documentation. Updated November 16, 2026

MountDev Ops Integrations documentation | Updated November 16, 2026

Our Commitment

MountDev Ops is built security first. Your business data belongs to you, and you can export all of it at any time. We never use analytics or tracking of any kind inside MountDev Ops.

Every update passes all 5,581 automated checks before it ships, and every screen change is walked through before release.

What They Are

A connection you added yourself in Settings, Connections (an address and a key for any outside service that has an API) can carry Actions. An Action is one call to that service, described once: how to call it, what to ask you for, and what to keep from the answer. Each turned on Action shows up in the workflow builder as a step named after the connection, such as example_hosting.create_site, so a workflow can use it like any other step.

Nothing here is made for one service. Any business can describe any service this way, and its own AI agent can help.

What an Action Holds

  • A name and a label. The name is lower case letters, digits, and _ (like create_site). The label says in plain words what it does, such as "Create a site".
  • How to call it. Get, Post, Put, Patch, or Delete, and a path with slots, like /sites/{site}. The path stays on the connection's own address. Every slot is an input.
  • Inputs. Each is text, a number, yes or no, or a choice. An input can be required, and can be Private: a Private input is never shown in a check, an error, or the Activity Log, and it must come from an earlier step, never typed into the workflow (a workflow with a typed one is not saved).
  • A body. Sent as JSON or as form fields. Form fields may use names with brackets, such as meta[key], sent as written. A GET call has no body. An Action can also always send up to 10 fixed fields (a name and a short value, at most 100 characters); they are shown on the approval card.
  • Values to keep. Named values taken from the answer by where they are, like data.site.id, as text (up to 2,000 characters), a number, or yes or no. They go on to the steps after this one. A value that looks like a key is not kept, and the step says so (not_kept). A kept value cannot be named like a key, or taken from a place named like one (a name or path with key, secret, token, password, or credential). The run history keeps only numbers, yes or no values, and ids, never text.
  • Waiting for a job. Some services answer a slow request with a job number. An Action can say where the job's status is, which values mean done or failed, how often to ask (every 60 to 3600 seconds), and how many times at most (up to 120). The call is made once, then the workflow waits and asks again until the job is done or failed, or it has asked as many times as allowed. Asking again does not use up the step's retries. A job that fails ends the step with "The service says the job failed." One that never finishes ends it with "The job did not finish after N checks."
  • A flag. None, Spends Money, Deletes Something, or Emails a Client. A flagged Action always stops for a person to confirm every run, and a workflow cannot make it automatic. A Delete call, and any path or name with the words delete, remove, restore, reset, disconnect, abort, deactivate, or suspend (also written together in camel case, like deleteSite), starts flagged Deletes Something, even a Get or a Post, because some services delete that way. A person can lower that flag, or change it to another kind, on the screen by giving a reason, which is kept. An AI agent can only keep the starting flag or make it stricter (from none to any flag), and can never lower it or change its kind.

States and Who Says Yes

Every Action is a Draft, On, or Off. A new one is a Draft. A Draft or Off Action is not offered to workflows. Turning one on needs an owner or admin who manages connections: when an AI agent asks, the answer is waiting_for_approval and the card in the workspace starts with the risk ("It only reads", "It changes things at the service and runs on its own", or the flag), then shows the address, method, path, every input, and every fixed body value. A yes covers exactly what the card showed: if the Action is changed after it was asked about, the yes turns nothing on and an owner must be asked again. Any change to an Action that was on puts it back to a Draft, so what was approved is what runs. An Action that is not flagged and changes something says plainly on that card that it runs on its own.

Removing an Action waits for a yes too, and is refused while a saved workflow uses it (the answer names the workflow).

For Your AI Agent (Through the MCP)

Everything here is also available through your account's MCP, with the same permission (integrations.manage, owners and admins) and the same approvals. Each tool is its own permission and level. Nothing here ever opens, shows, or accepts a key, and no tool makes the call itself.

ToolActionPermissionLevelWhat it does
connections_listintegrations.connections.listintegrations.connections_listReadThe connections you added: name, address, where the key goes (never the key), whether a key is saved, fixed address on or off, health, and how many actions each has
connections_draftintegrations.connections.draftintegrations.connection_draftWriteAdd a connection with no key, or change a connection's address or where its key goes. Waits for an owner or admin's yes. A change of address or key place removes the saved key, so a key is never sent somewhere new until a person types it again
connection_actions_listintegrations.actions.listintegrations.actions_listReadOne connection's Actions with state, flag, inputs, and what each keeps
connection_actions_scanintegrations.actions.scanintegrations.actions_scanReadRead an OpenAPI file (JSON text you give; nothing is fetched here) and list its calls with a suggested flag and the paths that cannot be read
connection_action_saveintegrations.actions.saveintegrations.action_saveWriteMake or change one Action, always as a Draft, with the flag rules above
connection_actions_importintegrations.actions.importintegrations.actions_importWriteMake Drafts from chosen calls of the same file (at most 50 a call; an Action that exists is skipped)
connection_action_checkintegrations.actions.checkintegrations.action_checkReadWhat an Action would send with the values you give, with Private inputs hidden. Sends nothing
connection_action_enableintegrations.actions.enableintegrations.action_enableWriteTurn off at once, or ask to turn on (waits for a yes)
connection_action_removeintegrations.actions.removeintegrations.action_removeDeleteAsk to remove an Action (waits for a yes)

An agent builds the workflows that use these Actions with the workflow tools (workflows_catalog lists each turned on Action with its flag, workflows_save always saves off, workflows_test, workflows_enable waits for a person). See Workflows for Your Assistant.

What no tool can do, and why:

  • Type or replace a key. A key never passes through an assistant. A person types it in Settings, Connections.
  • Make a real call or Check Now. A call opens the key. A person does that on the screen; an agent can run a workflow only after a person turns it on.
  • Turn on Send From a Fixed Address. It decides where the key travels. A person turns it on.
  • Clear a flag. Only a person decides that something that deletes, spends, or emails a client may run on its own.

Reading an OpenAPI File

Many services publish an OpenAPI 3 file. Give it as JSON text (YAML is not read) and the scan lists each call it can read, with a suggested name, label, inputs (including every slot in the path), whether the body is form fields or JSON, and a suggested flag. A path the scan cannot read, such as one with [ or + in it, is listed so a person can write that Action by hand. At most 300 calls and 2 MB at a time.

What a Call Can Reach

A value that fills a slot in the path is encoded, and . or .. is refused, so a value can never climb out of the connection's own address. The final address must stay under the connection's address and path. A connection's short name cannot be the name of a built-in step (core, github, storage, reports, sites, and the like); it gets a number added instead, and a built-in step always wins.

Where the Key Goes

The connection holds the key, sealed in your account, and says where it goes (a bearer token, a header you name, a user name and password, or the address). An Action never holds, shows, or returns it. A call goes only to the connection's own address, through the same safety checks as the Call a Web Service step: https only, nothing private, no redirects followed. When the connection's Send From a Fixed Address switch is on, every call goes only through the fixed address, never directly. A key that is not saved yet stops the step with a plain sentence.

What a Test Run Does

A workflow's Test Run checks the Action's inputs and says it would call. It sends nothing.

Screens

The Actions list and editor on a connection's page in Settings, Connections are planned (Planned). Until they ship, Actions are made and managed through the MCP tools above.

Who Can Use It

Owners, admins, and anyone whose role allows managing connections. A member cannot grant these tools to an AI agent.

Ask Us Anything

We’d love to hear from you!