MCP Tools

MCP (Model Context Protocol) is ThinkWork’s OSS path for external tools. MCP servers expose capabilities that the tenant platform agent can call during work, while ThinkWork controls registration, assignment, auth resolution, and audit.

Use MCP tools when the tenant platform agent needs to call remote capabilities during a turn, such as CRM lookups, ticket updates, internal API operations, search, or database queries exposed by an MCP server.

How it works

Admin registers MCP server
    -> Assigns it to a template or Space
    -> Invocation resolves tenant or user auth
    -> Runtime connects to MCP server
    -> Tools are discovered
    -> Tenant platform agent calls tools during the turn
    -> Calls are logged in the thread timeline

Registration

MCP servers are registered at the tenant level, then assigned to templates or Spaces depending on the product surface.

Typical config includes:

Field	Description
Name	Display name for the MCP server
URL	HTTP endpoint for the server
Transport	Streamable HTTP or SSE
Auth	None, tenant API key, or per-user OAuth

URL is read from the record on every call

The MCP endpoint URL lives in tenant_mcp_servers.url. The runtime resolves it on every outbound call. There are no hardcoded defaults and no env-var fallbacks for MCP base URLs. Rename the record in admin and the next request uses the new host.

Authentication

MCP tools commonly use one of three patterns:

None for private-network services or systems with external network controls.
Tenant API key for a shared secret resolved at invocation time.
Per-user OAuth when downstream tools need user-scoped access.

Runtime behavior

At runtime, ThinkWork discovers tools from the MCP server and adds them beside built-in tools and skill-provided tools. If an MCP server is unavailable, the worker can continue with its remaining toolset.

MCP usage is tracked in the thread timeline, including:

tool name
MCP server attribution
input and output previews
execution status