cliver-core includes an MCP client that supports stdio and streamable HTTP servers. Connect a client to an AgentCore and its tools become available to the model.

from cliver_core import AgentCore
from cliver_core.mcp import MCPClient

client = MCPClient.from_discovery()
agent = AgentCore(bindings, builtin_tools=[...], mcp_client=client)

Discovery

MCPClient.from_discovery() reads CLIver-local MCP configuration — ~/.cliver/mcp.json, project .cliver/mcp.json, and drop-in configuration files. Pass extra= to add servers programmatically.

Call agent.chat(..., mcp_servers=["server-name"]) to advertise only a subset for one call. The client starts lazily (and idempotently) before tools are gathered, so the first turn already sees MCP tools.

Tool Naming

MCP tools are advertised as mcp__<server>__<tool>. They share the same execution path as builtin tools — including permission categories and events. A server’s self-declared readOnlyHint annotation maps to the SAFE category; it is a hint, never a permission boundary. Override individual tools with safety_overrides=.

Prompts and Resources

MCP prompts and resources are read on demand — never registered as model tools and never listed at startup:

prompts = await client.list_prompts()
messages = await client.get_prompt("server-name", "args-prompt", {"city": "Paris"})

resources = await client.list_resources()
contents = await client.read_resource("server-name", "demo://document")

This keeps the core a pure reader: cataloguing, formatting, caps, and whether to send MCP content to the model are application concerns.