Skip to content

MCP Integration

Kombuse implements an MCP (Model Context Protocol) server that exposes project data — tickets, comments, agents, and the database — as tools that any compatible AI client can use. This guide explains what MCP is, how to enable it for different AI clients, and what tools are available.

MCP (Model Context Protocol) is an open standard that allows AI assistants to interact with external tools and data sources through a structured interface. Rather than copying and pasting context into a chat window, an MCP-connected AI client can read tickets, post comments, and query the database directly.

Kombuse runs an MCP endpoint at /mcp on its local HTTP server (server name: kombuse, version 0.1.0). A lightweight bridge process (mcp-bridge.mjs) translates between the stdio transport that MCP clients expect and that HTTP endpoint. When the Kombuse desktop app starts, it writes the server port to ~/.kombuse/server-port; the bridge watches this file and forwards JSON-RPC messages to the running server.

Any MCP-compatible tool — Claude Code, Claude Desktop, Cursor, or others — can connect to Kombuse by pointing to the bridge process.

The Settings page includes three MCP-related cards in the Chat tab. These automate configuration for Claude Code and Codex specifically.

The MCP integration toggles in the Chat tab of Settings

The three cards are:

  1. Codex — the “Enable MCP for Codex” toggle writes an MCP server entry to ~/.codex/config.toml. Toggling this setting stops all active Codex backends.

  2. Claude Code — the “Enable MCP for Claude Code” toggle writes an MCP server entry to ~/.claude/settings.local.json. This applies globally to all Claude Code sessions. Toggling this setting stops all active Claude Code backends.

  3. MCP Security — the “Allow anonymous write access” toggle controls whether external MCP connections can create and modify data. See the Anonymous Write Access section for details.

For other MCP clients (Claude Desktop, Cursor), manual configuration is required — see the next section.

All MCP clients connect to Kombuse through the same bridge process. The configuration tells each client how to launch the bridge.

Claude Code — the easiest setup is the Initialize Project button on the project settings page. The button description reads: “Set up AGENTS.md, MCP config, and .kombuse directory for this project.” Clicking it creates a .mcp.json file in the project directory with the correct bridge path.

The Initialize Project section in project settings

The generated .mcp.json has this structure (the bridge path reflects the actual Kombuse installation location on the system):

{
"mcpServers": {
"kombuse": {
"type": "stdio",
"command": "node",
"args": ["/path/to/kombuse/server/mcp-bridge.mjs"]
}
}
}

Claude Desktop — the same mcpServers block is added to the Claude Desktop configuration file (claude_desktop_config.json). The bridge command and args are identical to the Claude Code configuration.

Cursor — the server entry is added to .cursor/mcp.json in the project directory, or configured through Cursor’s MCP settings UI. The format is the same as the Claude Code configuration above.

By default, external MCP connections are read-only. Read tools such as list_tickets, search_tickets, get_ticket, query_db, list_agents, list_tables, and list_api_endpoints work without any special permissions.

Write operations — create_ticket, add_comment, update_ticket, update_comment, create_agent, update_agent — are blocked unless anonymous write access is enabled.

The “Allow anonymous write access” toggle is in Settings → Chat tab → MCP Security. When enabled, external tools connected via MCP can create and modify tickets, comments, and agents without authentication. When disabled, MCP connections are read-only. The setting defaults to disabled.

Enable it when AI agents need to create tickets, post comments, or manage agents through MCP. Disable it to prevent unintended modifications from external tools.

Kombuse exposes a comprehensive set of MCP tools across six categories. Desktop tools are only available when running the desktop app.

ToolDescription
create_ticketCreate a new ticket with title, description, and optional fields
get_ticketRead a ticket with comments, labels, and metadata (25 KB response cap)
list_ticketsList tickets with filtering by status, assignee, labels, and sorting
search_ticketsFull-text search across ticket titles, bodies, and comments
update_ticketUpdate status, priority, labels, assignee, title, or body
list_projectsList all projects in the workspace
list_labelsList all labels available in a project
ToolDescription
add_commentAdd a comment to a ticket (supports @profile and #ticket mentions)
get_ticket_commentRead a single comment by ID
update_commentEdit a previously posted comment
ToolDescription
list_agentsList agents with optional project and enabled/disabled filters
create_agentCreate a new agent with system prompt, permissions, and config
update_agentUpdate an agent’s prompt, permissions, config, or enabled state
ToolDescription
query_dbExecute read-only SQL queries (SELECT only, max 500 rows)
list_tablesList all tables and views in the database
describe_tableShow the schema of a specific table with columns and foreign keys
ToolDescription
list_api_endpointsDiscover all available Kombuse REST API endpoints
call_apiCall a GET endpoint on the Kombuse API

These tools are only available when Kombuse is running as a desktop application.

ToolDescription
list_windowsList all open Kombuse desktop windows
open_windowOpen a new desktop window, optionally navigating to a path
navigate_toNavigate an existing window to a new path
execute_jsExecute JavaScript in an isolated desktop window
wait_forWait for a CSS selector to appear in a window’s DOM
take_screenshotCapture a screenshot of a window as PNG
save_screenshotCapture and save a screenshot to disk
close_windowClose a desktop window