Connect an AI agent over MCP

Let your own AI agent build warehouse processes for you. Picked runs an MCP server for its process builder: an agent can read the building blocks, draft a process, test it, and publish it, under the same checks as a person in the builder.

When to use this

Use this recipe when you work with an MCP-capable agent, such as Claude Code, and want it to set up or change processes. It isn't a way to push orders or read stock: those go through the REST API, as in Push orders, with its own tokens.

You need a Professional or Enterprise plan, and the permission to edit processes in your workspace.

Get a token

Open Settings → AI agent access in Picked and issue a token. Give it a name that says which agent uses it. The token works only for this workspace and only for the process-builder tools. Picked shows the secret once: store it in your secret manager, never in code or chat.

An API token from Settings → API tokens doesn't work here, and an agent token doesn't work on the REST API.

Connect the agent

The quickest way: on the same page, copy the line under Onboard your agent and paste it into your agent. It reads:

Fetch https://picked.nl/mcp/prompt.md

The agent fetches Picked's instructions and sets itself up, asking you for the token.

To connect a client yourself:


Endpoint	`https://picked.nl/mcp/process-builder`
Transport	Streamable HTTP
Authentication	`Authorization: Bearer` followed by the token
Rate limit	Per token: 60 calls a minute on Professional, 240 on Enterprise

For example, in Claude Code:

claude mcp add --transport http picked https://picked.nl/mcp/process-builder \
  --header "Authorization: Bearer $PICKED_MCP_TOKEN"

What the agent can do

Tool	What it does
`list_nodes`	Lists every building block of a process, with the settings each one takes.
`get_node_schema`	Describes one building block.
`create_process`	Drafts a new process.
`edit_process`	Changes a draft.
`simulate_process`	Test-runs a draft without touching the warehouse.
`publish_version`	Makes a tested draft the live version.
`rollback_version`	Brings back a version that was live before.
`query_event_store`	Reads your workspace's event history, to see what happened on the floor.

The agent works in a loop: read the building blocks, draft, simulate, repair from what the checks say, and publish.

What keeps your warehouse safe

Drafts first. Creating and editing never change the live process. Only publish_version and rollback_version do, and a rollback only brings back a version that was live before.
Test before publish. publish_version needs a recent successful simulation of exactly that draft, by the same token.
The same checks as in the builder. An agent can't save a process a person couldn't.
Everything is recorded. Every tool call is logged in your workspace under the token that made it.
You stay in control. Revoking the token on Settings → AI agent access stops the agent at once.

Errors to handle

Status	What to do
`401`	The token is missing, wrong or revoked.
`403`	The token isn't an agent token, or the workspace isn't on Professional or Enterprise.
`429`	Wait the seconds in `Retry-After`.

A tool that refuses a draft answers with the reason in its result. A good agent reads it and repairs the draft.