What is Cortex Hub

Cortex Hub is a backend you host yourself. Your AI coding agents connect to it over one MCP endpoint, and through it they share memory, team knowledge, a code graph of every indexed repo, quality gates and a task queue.

   Claude Code   Cursor   Windsurf   VS Code   Antigravity   Codex
        │           │         │          │           │          │
        └───────────┴─────────┴────┬─────┴───────────┴──────────┘
                                   │  MCP over HTTP, API key
                         ┌─────────▼──────────┐
                         │   cortex-mcp       │  32 tools
                         ├────────────────────┤
                         │   cortex-api       │  REST, dashboard, Conductor
                         │   Qdrant           │  vectors
                         │   GitNexus         │  AST code graph
                         │   SQLite (WAL)     │  sessions, tasks, knowledge
                         │   Ollama           │  local embeddings
                         │   llm-proxy        │  LLM gateway
                         └────────────────────┘

What an agent gets

Area	Tools	Read more
Sessions	cortex_session_start, cortex_session_end, cortex_changes	Sessions and recall
Code	cortex_code_search, cortex_code_context, cortex_code_impact, cortex_detect_changes, cortex_cypher, cortex_list_repos, cortex_code_reindex	Code intelligence
Knowledge and memory	cortex_knowledge_*, cortex_memory_*, cortex_file_notes	Knowledge and memory
Quality	cortex_quality_report, cortex_plan_quality, cortex_tool_stats	Quality gates
Orchestration	cortex_task_*, cortex_agents	Conductor

All 32 are listed with their parameters in the MCP tools reference.

What runs on your machine

The hub alone is passive: an agent has to call it. What makes agents call it reliably is the part installed into each project by scripts/install.sh:

• MCP config for each IDE it finds, pointing at your hub.
• /cs and /ce, the slash commands that open and close a session. See Slash commands.
• Eight Claude Code hooks that block edits until the session has recalled what the team knows, and block commits until build, typecheck and lint pass. See Hooks.
• Rules files for Cursor, Windsurf, VS Code Copilot and Codex that tell those agents the same workflow.

For multi-agent work, a machine can also run a worker (cw), which takes tasks from the hub and runs them with a local engine (agy, claude, codex or gemini).

Where to go next

• Already have a hub URL and key? Quick start.
• Need a hub? Self-host it with Docker Compose.
• Want to see the orchestration? Open the Conductor showcase.

Requirements and cost

From the project README:

Resource	Minimum	Recommended
CPU	2 vCPU	4 vCPU
RAM	4 GB	8 GB
Disk	20 GB	50 GB
OS	Ubuntu 22.04+	Ubuntu 24.04 LTS (any Linux with Docker 24+)

All services run in Docker on your server. The README puts the cost at about $5 a month for a small VPS, plus whatever LLM usage you route through the gateway. Cortex Hub is MIT licensed.