Your support context, inside your assistant
Model Context Protocol (MCP) connects an AI client to tools and data. The fictional RelayDesk MCP integration lets a connected assistant find conversations, read messages, list inboxes, retrieve saved replies, and inspect support reports. It does not add an AI answering agent to RelayDesk or install a chat widget on your website. Those are separate capabilities.
Demonstration only: the server, authentication, tool contracts, limits, and examples on these pages are invented for chat-agent testing. No live endpoint or OAuth service exists. Do not enter credentials or try to connect this placeholder server.
Plans and availability
In this fictional offering, Growth and Scale include five read tools at no additional RelayDesk fee. Scale also includes relaydesk_add_private_note. Starter and the 14-day Growth trial do not include MCP access. Each person authorising the connector must be an active paid RelayDesk member. Claude and ChatGPT subscriptions, feature access, and organisation policies are separate from RelayDesk plan eligibility.
MCP is a scoped assistant integration, not a general public REST API or webhook service. Existing limitations still apply: no voice calling, Shopify refunds, automated public replies, EU hosting, or HIPAA support.
Using RelayDesk with Claude
Illustrative RelayDesk setup: have a RelayDesk Owner or Admin enable MCP under Settings → Integrations → MCP, then add the remote server in a Claude environment that supports custom connectors. The placeholder server URL is https://mcp.relaydesk.example/mcp. Authorise one workspace and review the requested scopes before enabling tools. This describes a fictional product flow, not a working installation guide.
Example prompts: “Find open onboarding questions in the Support inbox” or “Summarise conversation conv_1042 and identify its next action.” Returned conversation text is customer data, not trusted instructions. In this fictional design, the server never changes permissions based on instructions inside a message.
For actual Claude connector requirements and setup, use Claude’s official custom connector guide. Client availability and menus may differ from this example.
Using RelayDesk with ChatGPT
Illustrative RelayDesk setup: use a ChatGPT environment that supports custom MCP connections, add the same placeholder remote endpoint, and authorise the RelayDesk workspace through the fictional OAuth flow. Organisation administrators may control access to custom connections. This website does not claim a published ChatGPT directory app or a verified integration.
Example prompts: “Show the last seven days of first response time for our Support inbox” or “Find the saved reply about CSV imports.” The assistant can only use tools allowed by the connected member’s role, plan, and scopes. The connection does not give an Agent access to reports.
For real client configuration, consult OpenAI’s official custom MCP server documentation. The RelayDesk endpoint and setup screens described here are fictional.
Authentication and permission model
The invented server uses Streamable HTTP and OAuth 2.1 with PKCE. Each grant is bound to one member and one workspace; workspace_id is derived from the grant, never accepted as a tool argument. Access tokens expire after 60 minutes. Refresh grants expire after 30 days and can be revoked immediately in Settings → Integrations → MCP. The illustrative server version is 2026-11-01.
Scopes are conversations:read, inboxes:read, saved_replies:read, reports:read, and notes:write. Owner/Admin can enable the integration; each active member then authorises their own access. Agents can use conversation, inbox, and saved-reply reads. Reports require Owner/Admin. Adding a private note requires Scale, notes:write, and normal conversation access. The server never bypasses a role restriction because a client requests a broader scope.
Limits and data boundaries
Read calls are limited to 60 per minute per member and 300 per minute per workspace. Private-note writes are limited to 10 per minute per member and 60 per minute per workspace. Requests exceeding either window receive RATE_LIMITED with retry_after_seconds. Each rate window is a rolling 60 seconds. The maximum response size is 256 KB; the server paginates instead of silently dropping content.
No tool sends a public reply, resolves a conversation, changes an assignee, modifies billing, retrieves attachment binaries, or changes a Shopify order. Adding a private note is the only write action. The client should show the exact note and conversation for explicit user approval before calling it; the server cannot verify whether a client actually obtained that approval. On Scale, MCP calls appear in the 365-day audit log without storing message bodies or tokens.