MCP servers
MCP (the Model Context Protocol) is an open standard for exposing tools to AI models. Connect an MCP server to RoutMe and your assistants can use its tools: look up an order, fetch a customer's plan, check stock.
Two parts of RoutMe use them:
- the chat widget assistant, which only ever uses read-only tools, and
- Agent nodes in flows, which use the servers you attach to the node.
Connect a server
Open MCP servers
In the dashboard choose MCP servers, then Add connection.
Enter the details
Give it a name your team will recognise, like "Acme CRM", and the server URL, such as
https://mcp.acme.com. Add an auth token if the server needs one.Test it
Choose Test connection on the connection. RoutMe connects, performs the MCP handshake and reports the server's name, version and the tools it exposes. If it can't connect, the message says why.
Choose where it's used
Pick which widgets may call it, and attach it to Agent nodes in your flows.
The URL must start with http:// or https://; use https for anything but local testing. If you
set an auth token, RoutMe sends it as Authorization: Bearer <token> with every call to that server.
A connection can be active or disabled. A disabled connection is skipped by the chat assistant, so you can pause it without deleting it.
Use servers in the chat widget
On a widget, Edit widget has an MCP servers setting:
- All servers (default)Every active server in the app, including ones you connect later.
- Only selectedJust the servers you tick for this widget.
- Selected, but none tickedThe widget has no external tools at all.
This lets a support widget reach your order system while a marketing-site widget reaches nothing. See Shape the assistant for the rest of the per-widget settings.
Your tools must be marked read-only
The chat assistant talks to anonymous website visitors, so it must never change anything. RoutMe
enforces that on the tools themselves, not by asking the model to behave: a tool is offered to the
assistant only if the server marks it read-only, using the standard MCP tool annotation
readOnlyHint: true.
{
"name": "get_order",
"description": "Look up an order by its number.",
"inputSchema": {
"type": "object",
"properties": { "orderId": { "type": "string" } },
"required": ["orderId"]
},
"annotations": { "readOnlyHint": true }
}
A tool with no annotation, or with readOnlyHint: false, is left out. The assistant can only look
things up and show them, so it never gets a tool that might change data. If every tool on a server
is missing the annotation, the assistant has no tools from that connection and will tell visitors
it has no access. If the assistant can't see a tool you expected, check this first. RoutMe logs a
warning naming the connection and the skipped tools in this case.
Flow Agent nodes do not apply this filter: they can call every tool on the servers attached to the node. Attach only servers whose tools you are happy for the node to use.
How a call appears to the visitor
While a tool runs, the widget receives tool-call events and can show "checking your order" instead
of a spinner. The assistant sees each tool as mcp_<server>_<tool>, built from the connection name
and the tool name, so a server called "Acme CRM" exposing get_order appears as
mcp_acme_crm_get_order.
Build a server RoutMe can use
RoutMe speaks MCP's Streamable HTTP transport: a JSON-RPC 2.0 request and response per call. A server can answer with a single JSON body or a one-shot event stream carrying that same response.
- Name
Protocol version- Type
- 2025-03-26
- Description
The version RoutMe sends in the
initializehandshake.
- Name
Calls- Type
- initialize, tools/list, tools/call
- Description
What RoutMe uses: the handshake (followed by
notifications/initialized), tool discovery and tool calls.
- Name
Timeout- Type
- 20 seconds
- Description
Per request. Keep tools fast; the visitor is waiting.
- Name
Not supported- Description
Resumable streams, server-initiated push and batched requests.
Keep tools small and fast, return only what's needed to answer, and describe each tool and its parameters clearly: the descriptions are what the assistant reads when it decides which tool to call.
Reliability
- If one server is down or fails the handshake, RoutMe skips it and carries on with the others, so a flaky integration doesn't take down the whole widget.
- A single assistant turn allows up to six rounds of tool calls before it must answer.
- A tool call that fails comes back to the assistant as an error, so it can tell the visitor it couldn't look that up instead of guessing.
Frequently asked questions
What is an MCP server?
A Model Context Protocol server exposes tools, such as looking up an order, that an AI assistant can call. You connect your own server to RoutMe with its URL and an optional bearer token.
Can the chat assistant change data through an MCP server?
No. The chat assistant is only offered tools that the server marks read-only with the readOnlyHint annotation. Tools without that annotation are not exposed to it.
Why can the assistant not see one of my tools?
The most common reason is a missing readOnlyHint: true annotation on the tool. Also check that the connection is active, that the widget has it selected, and that Test connection lists the tool.
How is my server authenticated?
If you add an auth token to the connection, RoutMe sends it as an Authorization Bearer header with every call to your server.
Can different widgets use different MCP servers?
Yes. Each widget uses all active servers or only the ones you select. Selecting none means the widget has no external tools.
Do flows use MCP servers too?
Yes. Agent nodes in flows can call the tools of the MCP servers attached to the node. Unlike the chat widget they are not limited to read-only tools.