Skip to content

MCP servers & Toolsets

Building an MCP server exposes your tools to DioscHub; this page is the admin side of that — registering the server, deciding which Assistants may reach it, and keeping a large tool catalog manageable. For the server-side contract your MCP server must satisfy, see MCP server development.

Register your server as an MCP instance — a name and its Streamable HTTP URL — as described in The MCP contract. Registering it makes DioscHub aware of the server; it does not yet expose its tools to anyone. If you later change the server’s URL (or OAuth token endpoint) to another origin while it has a stored credential, you must enter the credential again in the same save; see Register your server.

A server’s tools become reachable only once you attach the instance to an Assistant. That attachment is the outer boundary on what the Assistant can ever call: no Role can invoke a tool from a server that is not attached to its Assistant. Within that boundary, each Role’s tool scope decides which of the attached tools that Role’s users may actually call — see Assistants & Roles.

A server with a handful of tools needs nothing more — every allowed tool is handed to the model directly. A server (or several) exposing dozens of tools is a different problem: putting the whole catalog in front of the model at once dilutes its choices and burns context. A Toolset solves that. It is a named group of related tools that the assistant loads on demand, keeping only the relevant ones in play at any moment.

Toolsets are about managing scale and context — not about permissions. What a Role may call is still governed by the Assistant’s attached servers and the Role’s tool scope. A Toolset only changes when an allowed tool is presented to the model, never whether it is allowed.

Create a Toolset (POST /admin/toolsets) with a name and description, then assign tools to it. Assignment is per MCP instance — you map each of that instance’s tools to the Toolsets it belongs to (PUT /admin/mcp-instances/:name/toolset-assignments). A tool can belong to more than one Toolset; a tool you assign to none stays available in the instance’s default group, so nothing is ever stranded.

The Tools tab of an MCP instance showing 114 discovered tools, each with a per-tool Toolset assignment column, plus Refresh, New toolset, and Suggested toolsets controls.

The Tools tab of a registered server — here a 114-tool HR adapter. Each tool has a Toolset-assignment control; Refresh re-discovers the catalog, and Suggest toolsets proposes groupings for the unassigned tools.

  • A pinned Toolset is loaded up front — its tools are always in front of the model. Pin the small set of tools the assistant needs on nearly every turn.
  • An unpinned Toolset loads only when the conversation calls for it. The assistant brings a Toolset in when the task needs it and sets it aside afterward, so the model sees a focused set rather than everything at once.

Wherever you name a tool — in a Role’s allowlist, in an approval rule — use the form DioscHub expects. Across an Assistant the model sees each tool under its qualified name, <mcpInstanceName>_<toolName>, joined by a single underscore. Per-tool configuration keyed to one instance (Toolset assignment, approval rules) uses the bare tool name. The full rule is in Report tools by their bare name.

The assistant has a built-in tool named get_my_usage that reports the user’s own quota and usage (see Usage questions). If an MCP tool in the assistant’s reach resolves to the same name (for example an instance named get_my with a tool named usage), DioscHub withholds the built-in for that assistant and your tool keeps the name. The assistant then has no built-in way to answer quota questions, and the instruction that points it at get_my_usage is left out of the prompt. A warning is logged on every turn that finds the clash, naming the tool. Rename the MCP tool to get the built-in back. The same rule applies to report_problem, see Problem reports.

The Toolset name usage is taken by the built-in Toolset. A Toolset of yours called usage still works and is loaded by name as usual, but the built-in stays pinned next to it, so pick another name to avoid confusion.


Next: Models.