MCP

The Model Context Protocol (MCP) server can be enabled on the knot server by adding the following configuration to the knot.toml file:

    
[server.mcp]
enabled = true

This configuration enables the MCP server at the path /mcp. Notably, it does not require knot to have access to a Large Language Model (LLM).

Connecting a Client

The knot MCP server operates over HTTP transport and is accessible at the URL: https://<your-knot-domain.com>/mcp

For example, if your knot installation is hosted at knot.getknot.dev, the MCP server URL would be: https://knot.getknot.dev/mcp

OAuth2-Supported Clients

If your client supports OAuth2, it will connect to the MCP server and allow you to log in using your standard credentials. During this process, an API token will be generated for use.

Non-OAuth2 Clients

For clients that do not support OAuth2, you will need to manually generate a new API token via the web interface and provide it to the MCP client. Use the following format:

Key	Value
Authorization	Bearer `your token`

In both cases, the MCP client will have the same level of access to tools as the user who generated the API token.

The web assistant may ask the browser user to approve write-capable tool calls before they run. This approval prompt is part of the web chat flow only; external MCP clients connected to /mcp are not prompted by Knot’s browser confirmation UI.

Tool Modes

Knot’s MCP server supports two different tool modes, configured with the native_tools setting in your knot.toml:

Tool Discovery Mode (Default)

In tool discovery mode, tools are discovered on-demand to minimize context usage. This is efficient for AI assistants that need to manage their token usage.

Workflow:

Use tool_search(query="<operation>") to find the appropriate tool
Use execute_tool(name="<tool_name>", arguments={...}) to execute the tool

Example:

tool_search("list spaces") → Returns tool information
execute_tool("list_spaces", {}) → Executes the tool

Benefits:

Dynamic tool availability based on user permissions
Reduced memory footprint and context usage
Permission-based filtering
Request-scoped authentication

Native Tools Mode

In native tools mode, all tools are pre-loaded and directly available without discovery. This is simpler for clients that don’t support the discovery pattern.

Configuration:

    
[server.mcp]
enabled = true
native_tools = true

Workflow:

All tools are available directly in the tool list
Use tools by name with their required arguments
No discovery pattern needed

Example:

list_spaces() → Direct tool call
create_space({...}) → Direct tool call

Tools

The knot MCP server exposes built-in tools for managing spaces, templates, stack definitions, stacks, files, commands, and skills. Write-capable tools require approval when called from the web assistant; read-only tools run without a confirmation. External MCP clients connected to /mcp are not prompted.

For the full list grouped by what each tool operates on — plus whether each is native or on-demand — see MCP Tools.

Remote MCP Servers

Knot can connect to external MCP servers and expose their tools alongside the built-in tools. This allows you to create a unified interface for AI assistants to access tools from multiple sources.

Remote tools are namespaced with a prefix (e.g., ai.generate-text) to avoid conflicts with local tools.

For detailed configuration and usage, see Remote MCP Servers.

Navigation

MCP