How MCP tools work in SentientOne

Model Context Protocol (MCP) servers expose your systems as tools: look up an order, check stock, open a ticket. Add an MCP server to an agent on its Tools tab, and the agent can call those tools while it answers, through every channel: the API, the widget, the Playground and workflow Agent blocks.

The tool loop

1. The model gets the question plus the name, description and input schema of every tool on the agent's enabled MCP servers. Agents with knowledge also get a built-in search_documents tool.
2. If the model decides a tool would help, SentientOne calls your server with the arguments the model chose, and streams a tool_call event to you.
3. The result goes back to the model (cut to 2,500 characters) and to you as a tool_result event.
4. The model either calls another tool or writes the answer. This repeats for up to 8 rounds, a fixed limit. If the rounds run out, the model is made to answer with what it has.

If the model calls the same tool with the same arguments twice, the repeat is skipped and the model is asked to answer, so a confused model can't loop.

Limits

Remote servers only

SentientOne runs in the cloud, so it can only reach MCP servers on a public internet address over HTTP. Local servers started as a process (stdio) can't be called from the cloud: the Tools tab offers a stdio option, but servers added that way aren't discovered or called. See Connect an MCP server.

SentientOne is an MCP client. It doesn't run an MCP server of its own for other assistants to connect to; to call your agents from code, use the REST API.