home / mcp / cortex mcp server
Provides multi-level reasoning with in-memory sessions, async tasks, and real-time reasoning traces.
Configuration
View docs{
"mcpServers": {
"j0hanz-cortex-mcp": {
"command": "npx",
"args": [
"-y",
"cortex-mcp@latest"
]
}
}
}Cortex MCP is a self-contained reasoning engine that performs multi-level analysis on complex queries, maintains session state for follow-up questions, and supports asynchronous task execution for long-running reasoning tasks, all running locally in memory.
You run the Cortex MCP server locally and connect to it with an MCP client. Use the standard input/output transport to send queries and receive results. You can perform reasoning at different depth levels, resume sessions by an identifier, and receive progress updates for long-running tasks.
Prerequisites you need before starting include Node.js version 24 or newer and npm for package management.
Install and run the server directly without a formal installation step by using the following command. This downloads and launches the latest Cortex MCP server in one go.
npx -y cortex-mcp@latestIf you prefer to build from source or run the server as a long-running process, you can clone the project, install dependencies, build, and then start the server. The following commands outline that workflow.
git clone https://github.com/j0hanz/cortex-mcp.git
cd cortex-mcp
npm install
npm run build
node dist/index.jsCortex MCP exposes three reasoning depth levels with distinct token budgets and thought counts: basic, normal, and high. You select the level when you issue a reasoning request. The server maintains in-memory state for sessions, enabling you to continue conversations across multiple turns.
All reasoning happens locally in memory. The server validates inputs with strict schemas and uses standard error handling to avoid leaking internal details. Logs are written to standard error to keep the JSON-RPC stream clean on standard output.
You access the reasoning capability through a primary tool named reasoning.think. This tool validates inputs, selects the depth level, and returns a session identifier along with a trace of the thoughts generated during reasoning.
No output on standard output usually indicates a transport or client capture issue. Check that your client captures standard output while monitoring standard error for logs. If you need to inspect internal state or traces, use the available session resources and traces exposed by the server.
Async task support is available for long-running reasoning tasks, with progress notifications emitted as the task advances.
Performs multi-level reasoning on a query, validating inputs, applying the selected depth level, and returning a session identifier with thought traces.