MCP Marketplace

The MCP marketplace is where you give your agents real tools. Install a Model Context Protocol server once for your account, then attach it to the agents that should be able to use it.

> Three different "MCP" features — don't mix them up: > - MCP marketplace (this page) — install servers as *agent action tools* (create a ticket, look up an order). > - MCP as a knowledge source — index an MCP server's *content* into a website's knowledge base. See the MCP doc. > - Your site as an MCP server — let an external agent search *your* content. Also in the MCP doc.

How it fits together

Account (tenant)
  └─ installs an MCP server  ──────────────┐  (once, with auth)
                                           ▼
Website ─▶ Agent (Support / Sales / …) ─▶ linked MCP tools

You install once per account, then attach a server to one or more agents — which each belong to a website. That's how the same install can power tools on several of your sites without re-entering credentials.

Installing a server

In the dashboard, open MCP (the marketplace).

Browse the catalog (Slack, Linear, GitHub, Stripe, Notion, and more) and click Install.

Authenticate:

- Token servers — paste the API token. It's encrypted at rest (AES-GCM) and never shown again. - OAuth servers (e.g. GitHub, Stripe) — a popup walks you through authorization; no token to copy.

Use Test to confirm connectivity. Any error is surfaced on the install so you can fix and retry.

Linking a server to an agent

Open the website in the dashboard and go to the Agents tab.

Expand the agent you want to give tools to.

Under Tools (MCP servers), check the servers this agent may use.

The orchestrator merges the tools from every enabled server linked to the agent. Unchecking removes the link; the install itself stays for your other agents and sites.

Limits and safety

Call quotas keep tool use bounded: up to 10 calls per conversation and 100 per account per hour.

Write-confirm: any tool that changes something asks the visitor to approve a confirmation card before it runs. Read-only tools run automatically. See Agents → Safe actions.

Call log: every tool call is recorded (agent, tool, success, duration) so you can audit what your agents did.

SSRF protection: servers are reached outbound only; private/internal addresses are blocked.

Next steps

Agents — the agent team, handoff, and write-confirm

MCP — your site as a server, and MCP as a knowledge source