Quick start

You need a running hub and an API key. If you do not have a hub yet, self-host one first; it takes one docker compose up.

1. Get an API key

Open the dashboard, go to API Keys (/keys) and create a key. Give each person, and each worker machine, its own key: the hub tells sessions apart by key name, and Conductor notices are delivered per key.

2. Install into your project

Run this in the root of the project you work on. HUB_MCP_URL is your hub's MCP endpoint and HUB_API_KEY the key from step 1.

macOS / Linux

curl -fsSL "https://raw.githubusercontent.com/lktiep/cortex-hub/master/scripts/install.sh" \
  | HUB_MCP_URL=https://<your-mcp-host>/mcp HUB_API_KEY=<api-key> bash

Windows

$env:HUB_MCP_URL="https://<your-mcp-host>/mcp"; $env:HUB_API_KEY="<api-key>"
iwr -useb "https://raw.githubusercontent.com/lktiep/cortex-hub/master/scripts/install.ps1" -OutFile $env:TEMP\install.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File $env:TEMP\install.ps1

Set HUB_MCP_URL

Without HUB_MCP_URL, the installer falls back to a hub URL built into the script, which is not yours. If an IDE already has a cortex-hub MCP entry, the installer keeps it and does not write a new one.

The installer detects your IDEs (Claude Code, Gemini/Antigravity, Cursor, Windsurf, VS Code, Codex) and, for each, writes the MCP config and the workflow rules. For Claude Code it also installs the eight hooks and the /cs and /ce commands. It is safe to run again; --check reports what is installed without changing anything. Every flag is in the installer reference.

Once the installer has run once, Claude Code has a global /install command that runs it again in any project.

The key is also read from a HUB_API_KEY= line in the project's .env, or from an existing cortex-hub entry in another IDE's config.

3. Start a session

Restart your agent so it loads the new MCP server, then:

/cs

/cs opens a session on the hub, recalls the team's knowledge and your memory for this repo, shows who else is editing, and picks up any task assigned to you. In Claude Code, edits are blocked until it has run. See Sessions and recall.

When you are done:

/ce

/ce runs build, typecheck and lint, reports them to the hub, stores what you learned and closes the session with a summary the next session starts from.

Without the installer

Any MCP client can talk to the hub directly. For Claude Code:

claude mcp add --transport http --scope user cortex-hub https://<your-mcp-host>/mcp \
  --header "Authorization: Bearer <api-key>"

The same entry in ~/.claude.json needs "type": "http"; without it Claude Code reads the entry as a stdio server:

{
  "mcpServers": {
    "cortex-hub": {
      "type": "http",
      "url": "https://<your-mcp-host>/mcp",
      "headers": { "Authorization": "Bearer <api-key>" }
    }
  }
}

The dashboard's Settings page shows this snippet with your hub's URL filled in. Other IDEs are covered in Multi-IDE. You get the tools this way, but not the hooks that make an agent use them.

Next

• Self-host the hub
• Run a worker so other machines can take tasks
• cw reference