GZOO Cortex

Local-first knowledge graph for developers. Watches your project files, extracts entities and relationships using LLMs, and lets you query across all your projects in natural language.

“What architecture decisions have I made across projects?”

Cortex finds decisions from your READMEs, TypeScript files, config files, and conversation exports — then synthesizes an answer with source citations.

Why

You work on multiple projects. Decisions, patterns, and context are scattered across hundreds of files. You forget what you decided three months ago. You re-solve problems you already solved in another repo.

Cortex watches your project directories, extracts knowledge automatically, and gives it back to you when you need it.

What It Does

• Watches your project files (md, ts, js, py, json, yaml) for changes
• Extracts entities: decisions, patterns, components, dependencies, constraints, action items
• Infers relationships between entities across projects
• Detects contradictions when decisions conflict
• Queries in natural language with source citations
• Searches semantically — blends keyword and vector (embedding) similarity so queries match by meaning, not just keywords (optional; see Semantic Search)
• Routes intelligently between cloud and local LLMs
• Respects privacy — restricted projects never leave your machine
• Web dashboard with knowledge graph visualization, live feed, and query explorer
• MCP server for direct integration with Claude Code

How It Works

Cortex runs a pipeline on every file change:

1. Parse — file content is chunked by a language-aware parser (tree-sitter for code, remark for markdown)
2. Extract — LLM identifies entities (decisions, components, patterns, etc.)
3. Relate — LLM infers relationships between new and existing entities
4. Detect — contradictions and duplicates are flagged automatically
5. Store — entities, relationships, and vectors go into SQLite + LanceDB
6. Query — natural language queries search the graph and synthesize answers

All data stays local in ~/.cortex/. Only LLM API calls leave your machine (and never for restricted projects).

LLM Providers

Cortex is provider-agnostic. It supports:

• Anthropic Claude (Sonnet, Haiku) — via native Anthropic API
• Google Gemini — via OpenAI-compatible API
• DeepSeek (Reasoner, Chat) — strong reasoning, very affordable
• Groq — fast inference with free tier
• Any OpenAI-compatible API — OpenRouter, local proxies, etc.
• Ollama (Mistral, Llama, etc.) — fully local, no cloud required

Cost tracking uses provider-aware rates for DeepSeek, Gemini, Groq, and OpenRouter models — not a blanket Anthropic fallback.

Embeddings (for semantic search) are configured as a separate provider — independent of your chat model — so you can run chat on DeepSeek and embeddings on OpenAI. See Semantic Search.

Routing Modes

In cloud-first mode, all tasks route to your cloud provider. Ollama is not required and is only used if budget fallback is enabled. Hybrid mode routes high-volume tasks (entity extraction, ranking) to Ollama and reasoning-heavy tasks (relationship inference, queries) to your cloud provider.

Commands

Full CLI reference: docs/cli-reference.md

Web Dashboard

Run cortex serve to open a full web dashboard at http://localhost:3710 with:

• Dashboard Home — graph stats, recent activity, entity type breakdown
• Knowledge Graph — interactive D3-force graph with clustering, click to explore
• Live Feed — real-time file change and entity extraction events via WebSocket (from cortex serve only)
• Query Explorer — natural language queries with streaming responses
• Contradiction Resolver — review and resolve conflicting decisions

Remote Deployment

For access beyond localhost, bind to all interfaces and put Cortex behind a reverse proxy:

cortex serve --host 0.0.0.0

Example nginx config — protect /api/ and /ws with basic auth; serve static assets without auth (the dashboard injects the bearer token into HTML):

location /api/ {
    auth_basic "Cortex";
    auth_basic_user_file /etc/nginx/.htpasswd;
    proxy_pass http://127.0.0.1:3710;
    proxy_set_header Authorization "Bearer $CORTEX_TOKEN";
}

location /ws {
    auth_basic "Cortex";
    auth_basic_user_file /etc/nginx/.htpasswd;
    proxy_pass http://127.0.0.1:3710;
    proxy_http_version 1.1;
    proxy_set_header Upgrade $http_upgrade;
    proxy_set_header Connection "upgrade";
}

location / {
    auth_basic off;
    proxy_pass http://127.0.0.1:3710;
}

Set CORTEX_SERVER_AUTH_TOKEN or server.auth.token in config. When auth is enabled, Cortex injects the token into the dashboard HTML so API and WebSocket calls authenticate automatically.

MCP Server (Claude Code Integration)

Cortex includes an MCP server so Claude Code can query your knowledge graph directly:

claude mcp add cortex --scope user -- npx @gzoo/cortex mcp

This gives Claude Code 12 tools:

Architecture

Monorepo with eight packages:

• @cortex/core — types, EventBus, config loader, error classes
• @cortex/ingest — file parsers (tree-sitter + remark), chunker, watcher, pipeline
• @cortex/graph — SQLite store, LanceDB vectors, query engine
• @cortex/llm — Anthropic/Gemini/OpenAI-compatible/Ollama providers, router, prompts, cache
• @cortex/cli — Commander.js CLI
• @cortex/mcp — Model Context Protocol server (stdio transport, 12 tools)
• @cortex/server — Express REST API + WebSocket relay
• @cortex/web — React + Vite + D3 web dashboard

Architecture docs: docs/

Privacy & Security

• Files classified as restricted are never sent to cloud LLMs
• Sensitive files (.env, .pem, .key) are auto-detected and blocked
• API key secrets are scanned and redacted before any cloud transmission
• All data stored locally in ~/.cortex/ — nothing phones home

Full security architecture: docs/security.md

Built With

• SQLite via better-sqlite3 — entity and relationship storage
• LanceDB — vector embeddings for semantic search
• Anthropic Claude — cloud LLM provider
• Google Gemini — cloud LLM provider (via OpenAI-compatible API)
• DeepSeek — cloud LLM provider (reasoning + chat)
• Groq — fast cloud inference
• Ollama — local LLM inference
• tree-sitter — language-aware file parsing
• Chokidar — cross-platform file watching
• Commander.js — CLI framework
• React + Vite — web dashboard
• D3 — knowledge graph visualization

About

Built by GZOO — an AI-powered business automation platform.

Cortex started as an internal tool to maintain context across multiple client projects. We open-sourced it because every developer who works on more than one thing loses context, and we think this approach — automatic file watching + knowledge graph + natural language queries — is the right way to solve it.

Quick Start

1. Install

npm install -g @gzoo/cortex

If global install fails with EACCES, use a user prefix instead:

mkdir -p ~/.local
npm config set prefix ~/.local
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc
npm install -g @gzoo/cortex

Or install from source:

git clone https://github.com/gzoonet/cortex.git
cd cortex
npm install && npm run build && npm link

Verify: cortex --version (current release: 0.8.1)

2. Setup

Run the interactive wizard:

cortex init
cortex doctor                              # verify config, providers, and DB

This walks you through:

• LLM provider — Anthropic, Google Gemini, DeepSeek, Groq, OpenRouter, or Ollama (local)
• API key — saved securely to ~/.cortex/.env
• Routing mode — cloud-first, hybrid, local-first, or local-only
• Watch directories — which directories Cortex should monitor
• Budget limit — monthly LLM spend cap

cortex init writes global config to ~/.cortex/cortex.config.json. API keys go in ~/.cortex/.env.

3. Register Projects

cortex projects add my-app ~/projects/app
cortex projects add api ~/projects/api
cortex projects list                       # verify

4. Ingest, Watch & Query

Backfill existing files first — the watcher only picks up new changes:

cortex ingest "~/projects/app/src/**/*.ts"   # one-shot backfill
cortex serve                                 # dashboard + API + file watcher (recommended)

Don't run watch and serve together — they compete for file changes. The Live Feed shows real-time events from cortex serve only (file saves while the server is running).

cortex query "what caching strategies am I using?"
cortex query "what decisions have I made about authentication?"
cortex find "PostgreSQL" --expand 2
cortex contradictions

5. Web Dashboard

cortex serve                               # open http://localhost:3710

Remote access:

cortex serve --host 0.0.0.0

Auth is enforced automatically on non-localhost hosts. A bearer token is auto-generated and saved to ~/.cortex/.env (read it with grep CORTEX_SERVER_AUTH_TOKEN ~/.cortex/.env). Open the dashboard once with http://<host>:3710/?token=<token> — the token is embedded only for requests that already prove possession of it and is then kept for the browser tab (so anonymous visitors never receive it). API/WebSocket calls behind a reverse proxy use Authorization: Bearer <token>.

Excluding Files & Directories

Cortex ignores node_modules, dist, .git, and other common directories by default. To add more:

cortex config exclude add docs             # exclude a directory
cortex config exclude add "*.log"          # exclude by pattern
cortex config exclude list                 # see all excludes
cortex config exclude remove docs          # remove an exclude

Requirements

• Node.js 20+
• LLM API key for cloud modes — Anthropic, Google Gemini, DeepSeek, Groq, or any OpenAI-compatible provider
• Ollama — only for hybrid, local-first, or local-only modes (install)

Configuration

Config is layered — later sources override earlier ones:

API keys are stored separately in ~/.cortex/.env (never in config JSON).

cortex config list                       # see all non-default settings
cortex config set llm.mode hybrid        # switch routing mode
cortex config set llm.budget.monthlyLimitUsd 10  # set budget
cortex config exclude add vendor         # exclude a directory from watching
cortex privacy set ~/clients restricted  # mark directory as restricted
cortex doctor                            # validate setup

Full configuration reference: docs/configuration.md

Semantic Search (Embeddings)

Cortex blends keyword (full-text) search with vector similarity, so queries match by meaning rather than exact words. Embeddings are optional and off by default — enable them with a cloud embeddings provider (no local GPU or Ollama required):

cortex config set llm.embeddings.enabled true
cortex config set llm.embeddings.baseUrl https://api.openai.com/v1
cortex config set llm.embeddings.model text-embedding-3-small
cortex config set llm.embeddings.apiKeySource env:OPENAI_API_KEY
cortex config set llm.embeddings.dimensions 1536
# then add the key to ~/.cortex/.env:
echo 'OPENAI_API_KEY=sk-...' >> ~/.cortex/.env

The embeddings provider is independent of your chat provider — run chat on DeepSeek (or Anthropic, Groq, …) and embeddings on OpenAI. Any OpenAI-compatible embeddings endpoint works.

New files are embedded automatically as they're ingested. To build the index for a graph you already ingested, run a one-time reindex:

cortex reindex                 # all projects
cortex reindex my-app          # a single project

License

MIT — see LICENSE