Connect an assistant
ChatGPT, Claude, Cursor, and other assistants that speak MCP can read an environment and set up listeners and automations. They work through the same calls as the API, with the access you give them.
The address is https://pim-outpost.com/mcp.
Connect
- Claude. Open Settings, then Connectors, and add a custom connector with the address above. Sign in to Outpost when it asks.
- ChatGPT. Turn on developer mode in Settings, then Connectors. Create a connector with the address above and sign in. Some plans offer a connector read-only. If a tool that changes something is missing, check the plan.
- Cursor. Add a server in Settings, then MCP, with the address as its
url. Sign in when Cursor opens the browser. - Claude Code. Run
claude mcp add --transport http outpost https://pim-outpost.com/mcp, then sign in from/mcp. - Your own client. Send an Outpost key as
Authorization: Bearer $OUTPOST_KEY. The key's power decides which tools it sees.
Signing in opens a page in Outpost. It shows where the assistant sends you back to, then asks which environment it can reach and whether it may read or also manage. Read is the default. Disconnect an assistant in Settings, under Connected assistants.
What it can do
With read access an assistant sees the first group. With manage access it sees both.
Read
list_environmentsList the environments this connection can reach, with their organisation, Test or Production tier, and whether the API keys are saved.get_environmentRead one environment: whether the API keys are saved, its listeners with their automations, how many caught events are waiting or finished, and the latest finished events.list_eventsList caught events for one listener, 20 to a page, newest first by default.get_event_write_backRead the calls Outpost sent to the PIM for one caught event: method, path, and body.fetch_product_namesFetch the product name and number for up to 20 caught events, so a list can show names next to product ids.search_attributesSearch the environment's attribute definitions by name or number, and optionally by data type.list_category_attributesRead one page of category-level attributes: which attribute is assigned to which category, with the mandatory, locked, and propagate flags.list_assigned_productsCount the products that have one attribute, or list them 100 to a page.get_catalog_mapRead the saved map of which categories share product links.
Manage
list_assigneesList the users and user groups a task can be assigned to, from the PIM.create_listenerCreate a listener: one webhook endpoint for an environment.update_listenerRename a listener, or pause or resume it, or both.create_automationCreate an automation on a listener: a calculation on an integer or decimal attribute, a code on a text attribute, or a task from an attribute value or a name change.update_automationEdit an automation, or pause or resume it.replay_eventRun a skipped or failed caught event again through the automations that exist now.check_connectionCheck that the API keys saved on an environment still work.
An assistant cannot delete a listener, an automation, or events. It cannot create, list, or revoke an Outpost key, save API keys, or add an organisation. Those stay in Outpost. A webhook secret never enters the chat. After an assistant creates a listener, you open it in Outpost to issue the secret.
Keep it safe
- A caught event can carry text from your PIM or from a webhook. An assistant may be misled by it. Use read access to look around, and give manage access only to an environment where you accept that.
- Ask the assistant to say what it will change before it does. Most assistants ask you to approve a tool that changes something. That prompt belongs to the assistant, and Outpost cannot enforce it.
- A tool that takes longer than 40 seconds answers with a sentence instead of failing. A change may still have happened, so read the listener again before you repeat it.
- An assistant counts against the same limit as a key, about 120 requests a minute.