Many systems now expose their capabilities as an MCP server — a standard way for AI to call live tools: look up an order, create a ticket, check stock, book a job. Connect one to polite.ai and your agents can use its tools on real calls, on plans that include MCP connections (Pro and Scale).
The connection is held at organisation level and every call runs through polite.ai's own proxy: the server's credential is stored encrypted by polite.ai and attached on the way through, so it never appears in an agent, a prompt, or the Builder chat — and each agent only ever holds a scoped, revocable polite.ai credential.
Connect a server
In the Builder, ask for the tools you want ("the agent should look up orders in our CRM — here's our MCP server") and approve the Connect an MCP tool server panel it offers. Or connect ahead of time on Integrations, under MCP tool servers. Either way an organisation owner or admin completes it:
Give the server a name, a short slug (it prefixes the tool names agents
see), and its MCP URL — usually ending /mcp. polite.ai checks the server
before anything is saved.
Three kinds of server work: open (no sign-in), API key (paste the key; a custom header name is supported), and OAuth (a sign-in popup, exactly like connecting a calendar — polite.ai discovers the server's authorisation setup automatically, and only asks for a pre-registered client id in the rare case the provider requires one).
The server's tools are listed with their descriptions. All tools is the default — including tools the server adds later. Untick it to approve a fixed selection instead. Then pick which team members get them; the panel flags any member whose model can't use MCP tools.
The assistant then writes each member's guidance for the new tools — when to use them, and to confirm anything that creates or changes data with the caller first.
On live calls
Tools appear to the agent as slug_toolname — a server with slug crm and a
lookup_order tool gives the agent crm_lookup_order. Calls are proxied with
a time limit and a size cap, and a server that keeps failing is paused briefly
rather than slowing calls; answering a call never waits on the third-party
server.
Managing connections
Integrations → MCP tool servers shows each connection's health: its tools and when they were last synced, a Re-sync tools button that reports what the server added or removed, credential recovery (Update key for API-key servers, Re-authorise for OAuth), and Disconnect — which immediately revokes the agent credentials that referenced it. Per-agent, the agent editor's Tools & transfers tab lists the servers attached to that agent with a Detach for each; to change which tools are selected, re-run the connect step in the Builder.
Trust and data
A tool call sends its arguments — which can include what a caller just said — to the connected server, and the server's replies come back into the conversation. Connect servers you trust with caller data, treat their output as data rather than instructions, and prefer approving a fixed tool selection on servers you don't operate yourself.