# MCP for customer support > Peeve speaks the Model Context Protocol both directions: it serves your users' own AI agents your capabilities as typed tools over a scoped MCP endpoint, and it connects to an MCP server you run. Canonical: https://peeve.ai/mcp/ Peeve speaks MCP in both directions: it serves your capabilities to your users' own agents, and it connects to an MCP server you run. ## Serve your users' agents (inbound) Your user points their own agent (Claude, ChatGPT and other OpenAI agents, Hermes, OpenClaw, or any MCP client) at a scoped endpoint of the form `mcp.peeve.ai//u/` and gets your capabilities as typed tools, each with a stated effect, for example `change_plan`, `download_invoice`, `list_orders`. The agent gets the same permissions as that user, never more, and every grant is revocable and rate-limited (5 grants per user on Growth, unlimited on Scale). Anything that changes state is confirmed and written to a signed audit log; the model requests a capability by name and never holds a credential. Agent tool calls over MCP are logged but not billed today. ## Connect your MCP server (outbound) On Growth and above, paste your MCP server URL (and a token if it needs one). Peeve does a live handshake, discovers the tools it exposes, and you choose which the assistant may use. Anything that changes state runs through the same confirmation guardrail; credentials are encrypted per workspace and the model never sees a token. ## Machine-readable - Server card: https://peeve.ai/.well-known/mcp/server-card.json - Agent guide: https://peeve.ai/AGENTS.md - Markdown twins: append `.md` to any URL, or send `Accept: text/markdown`. ## More questions ### Which AI agents can connect to my product through Peeve over MCP? Claude, ChatGPT and other OpenAI agents, Hermes, OpenClaw, and any client that speaks MCP. The user points their agent at a scoped Peeve endpoint (mcp.peeve.ai//u/) and receives your capabilities as typed tools, each stating its effect before it runs. Nothing about the agent is special-cased. ### What is an MCP server for customer support? It is an endpoint that exposes your support and product capabilities to AI agents as typed tools. Peeve mints a scoped MCP endpoint per user, so an agent acting for that user can call your capabilities with stated effects, using the same registry that renders as a cursor for humans. ### How do I make my product usable by my customers' own AI agents? Expose your capabilities as tools an agent can call, with permissions and stated effects. Peeve does this from the same capability registry that powers its human-facing cursor: your user connects their Claude, ChatGPT, Hermes, OpenClaw or any MCP client to a scoped endpoint (mcp.peeve.ai//u/) and gets your capabilities as typed tools, with the same permissions the user has. ### Can an AI support agent connect to my own MCP server? Yes, on Peeve's Growth plan and above. Paste your MCP server URL and a token if needed; Peeve does a live handshake, discovers the tools it exposes, and you pick which the assistant may use. Anything that changes state runs through a confirmation guardrail and the model never sees a credential. ### Is it safe to let AI agents act in my product over MCP? Yes, when access is scoped and confirmed. With Peeve, agents inherit exactly the calling user's permissions and no more, grants are revocable and rate-limited, and any state-changing action is confirmed. The model requests a capability by name, so it cannot compose a raw request or reach a key.