Model Context Protocol (MCP) Integration
Verify client support, server permissions and deployment-specific setup before connecting tools.
MCP connects compatible AI clients with tool and data servers. An OpenAI-compatible inference endpoint alone is not an MCP server or a tool-execution permission system.
Confirm the client and server
Check that your selected client and the specific MCP server support a common transport and authentication method. Use the setup instructions for their installed versions.
This website does not install a universal THOX MCP configuration or a bundled set of filesystem, database and terminal servers on every device.
Limit the accessible resources
Review the exact tools and resources exposed by the server. Start with a narrow test directory or read-only test account where the integration supports it.
A server able to write files, query production data or execute shell commands has real privileges. Configure its access separately from model inference credentials.
Use the documented installation
Get the server from its verified maintainer and inspect the release instructions, license and requested permissions. Use the actual client configuration location rather than assuming a ~/.thox/mcp.json file.
Do not paste credentials into a shared example configuration. Use the credential mechanism supported by the client and server.
Test with synthetic data
Confirm the advertised tools, required permissions and processing location before a test. Use a harmless read operation against synthetic data first.
Review tool requests and outcomes. A generated tool call is a proposal until the client and authorized execution service permit and execute it; model confidence does not grant permission.
Understand website demo boundaries
The THOX website AI demos produce text or action proposals. They do not provide general MCP access to your filesystem, databases, shell or devices.
Consult the relevant workflow guide for its current capabilities and provider disclosure before choosing an integration.