purserDOCS
purserDOCS
HomePricing中文Sign in

Get started

purser docsGo live in 10 minutes

Channels

Website chatIn-app support pageEmailAmazon buyer messagesWhatsApp, Instagram and Messenger

AI support

Knowledge baseGlossaryHelp centerAI toolsHow the AI answersTry it and releasesTopics & QA

Inbox and team

InboxVoice messagesTeams and routingRules, macros and closingMarkets (several countries and currencies)Notifications and announcementsLearning centerCustomers

More

RefundsCompensation and order changesProduct feedback

Account and billing

Plans and billingAccount and data

Developers

Identity tokens: telling purser who the customer isREST API and Node SDKRefund endpoint (Actions API)WebhooksConnect your own AI assistant over MCP

AI tools

Give the AI your own API or MCP server: live stock, loyalty points, your own shipping data. Who the customer is comes from the system, and the AI only sees the fields you allow.

Out of the box the AI can use your knowledge base, store policies, synced products and a signed-in customer's orders. Data in your own systems, such as warehouse stock, loyalty points or tracking from your own carrier, can be handed to the AI as AI tools.

Find them under AI support → Tools. Only owners and admins can manage them, on the Growth plan and up: Growth allows up to 10 enabled tools, Scale has no limit.

Connections

Start by adding a connection, one of two kinds:

  • HTTP API: your API address, such as https://erp.example.com/api/.
  • MCP server: the address of an MCP server that supports Streamable HTTP. After adding it, click Pick tools from the server: purser lists the server's tools and you choose which to use.

Both take an authentication method: a Bearer token, a custom header, or none.

  • The credential is stored encrypted and never shown again, apart from its last few characters.
  • purser only calls the connection's own host, over https only, without following redirects, and never private network addresses.
  • Give purser its own read-only key, so you can revoke it on its own.

Tools

For each tool you set:

  • Tool name: what the AI calls it, in lowercase and underscores, such as check_stock.
  • When to use it: a description for the AI, such as "Real-time stock and restock date of a product in our warehouse; use when a customer asks whether something is in stock or when it's coming back".
  • Method and path (HTTP): GET puts the parameters in the query string, POST in a JSON body. The path can use {parameter}, such as /stock/{sku}.
  • Parameters, and who fills each one in:
Filled byMeaning
AIThe AI fills it from the conversation, such as a SKU or a size
Customer id, customer email (system)The system fills in the signed-in customer's details; the AI can't change them. For customers who aren't signed in, the AI doesn't see the tool
Customer's own order numberThe AI names an order, and the system checks it really is that signed-in customer's order; otherwise there's no call
Conversation id, customer's language, channel, marketFilled in by the system from the conversation
Fixed valueA constant you type, such as a warehouse code
  • Fields the customer may hear: one per line, dots for nesting and [] for lists, such as in_stock, eta.date, items[].name. The AI never sees fields that aren't listed, so it can't pass them on.

How the AI uses them

  • The AI can only call tools you turned on, and in this version all tools are read-only.
  • Each tool can be limited to certain kinds of questions (Offered for, like skills). With none selected it's offered in every conversation. At most 8 tools are offered per reply, so with many tools the AI stays fast and picks the right one more often.
  • When a question needs live data, the AI calls the right tool and answers from what it returns, with at most 3 calls per reply.
  • Like knowledge passages, what a tool returns goes through the pre-send check: numbers and statuses in the reply must come from the tool.
  • If a call fails, times out (5 seconds) or returns an error, the AI doesn't make up an answer: it hands the conversation to a person.

Test and publish

  • A saved tool is a draft: you can try it in Test chat right away, and customers get it once you publish, like skills.
  • Each tool has a Test panel that calls it once with the saved settings and shows the raw answer next to what the AI can see. For tools that need a customer, you can fill in a sample customer.
  • Recent calls lists each call's time, result, duration and error, kept for 90 days.

When an MCP tool changes

Once a day purser checks the MCP tools you turned on. If a tool's description or parameters changed on the server, or the tool is gone, it's turned off automatically and owners and admins are notified. Check it and turn it back on.

Help center

Publish articles from your knowledge base as a help center customers can browse: the same content as the AI, two levels of categories, the published version only, indexable by search engines.

How the AI answers

The AI answers only from your sources, every reply is checked before it is sent, and you decide per question type how far it may go on its own.

On this page

ConnectionsToolsHow the AI uses themTest and publishWhen an MCP tool changes