Model Context Protocol (MCP)#

Give your AI editor - Claude Code, Cursor, Codex - direct access to your Click Guide organization. An author can then say "populate my sources from this repository" and never leave the editor.

Click Guide is in development. The endpoint below is live in the backend and the console; the hosted URL follows when we ship.

What MCP is#

MCP is a small JSON-RPC 2.0 protocol that lets an AI editor discover and call tools running on a server. Click Guide exposes ten tools in v1:

Tool	Does
`list_applications`	List every application in your organization
`create_application`	Create a new application
`list_sources`	List documentation sources for one application
`add_source`	Ingest a markdown document as a new source
`remove_source`	Delete a source and every passage that came from it
`list_flows`	List every flow (draft + published) for one application
`save_flow`	Create or replace a flow draft (pass `flow_id` to update)
`publish_flow`	Freeze the current draft as an immutable version
`delete_flow`	Remove a flow and every version
`preview_answer`	Ask what the assistant would tell a user right now

Each tool is a thin wrapper over the same code the console uses. Nothing an MCP token can do is anything you could not do yourself in the console.

Get a token#

Sign in to the console.
Open the AI editor access (MCP) card.
Give the token a label - something like Claude Code on my laptop - and mint it.
Copy it immediately. It is shown once. Once you close the dialog, only the prefix remains; a lost token can be revoked from the same card but cannot be recovered.

Tokens start with cg_mcp_ and are valid for 90 days by default. They inherit your role: an author's token can add sources, an administrator's can also create applications.

Install it in your editor#

Claude Code#

Add a server to your project's .mcp.json (or your user config):

json

{  "mcpServers": {    "clickguide": {      "type": "http",      "url": "https://api.clickguide.co.za/mcp",      "headers": {        "Authorization": "Bearer cg_mcp_xxxxxxxxxxxx"      }    }  }}

Restart Claude Code. Every tool listed above appears under clickguide in the tool picker.

Claude Desktop (macOS / Windows)#

Claude Desktop reads MCP servers from a single JSON file. Location:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

Open it (create it if it does not exist) and add Click Guide alongside any servers already there. Claude Desktop currently ships stdio transport only, so wrap the HTTP endpoint with mcp-remote (a tiny bridge that ships with npx):

json

{  "mcpServers": {    "clickguide": {      "command": "npx",      "args": [        "-y",        "mcp-remote",        "https://api.clickguide.co.za/mcp",        "--header",        "Authorization: Bearer cg_mcp_xxxxxxxxxxxx"      ]    }  }}

Quit Claude Desktop fully (⌘Q on macOS, right-click tray → Quit on Windows - a window close is not enough) and reopen. In a new chat the Click Guide tools appear when you hit the tool-picker icon; try:

Use Click Guide to list my applications, pick the one called "My App", and show me its published flows.

If nothing shows up, open the developer console (Claude → Developer → Open Developer Tools) and look for mcp-remote in the log. The two usual causes are a stale bearer (rotate on the console's MCP card) or npx not being on PATH - install Node.js and try again.

Cursor#

Cursor reads MCP servers from ~/.cursor/mcp.json. Same shape as Claude Code:

json

{  "mcpServers": {    "clickguide": {      "url": "https://api.clickguide.co.za/mcp",      "headers": {        "Authorization": "Bearer cg_mcp_xxxxxxxxxxxx"      }    }  }}

Codex#

Codex CLI uses the --mcp flag pointing at an MCP config. Reuse either JSON above.

An example prompt#

Once the server is installed, this is enough to bootstrap an application from a codebase:

I have a Node.js project at ./. Read the README, the CHANGELOG, every .md under docs/, and every top-level route file. Create a Click Guide application called "My App" and add each document as a source using its path as the name. When you finish, call preview_answer with three questions a new user of my app would probably ask, and show me the answers.

Your editor will invoke create_application, then add_source for each file, then preview_answer three times, and hand you a summary. If any of your documentation does not answer the preview questions well, you know where to write more before your users see it.

Rules the token still follows#

Your role, not more. An author's token cannot create an application; a viewer's token cannot mint a token in the first place.
Your organization, not others. Every call is scoped to the organization you were signed in to when you minted the token.
Audited. Every call lands in your audit trail with your email as the actor, so a review can see what the agent did.

Revoke a token from the same console card the moment it is no longer needed.