Summary

It notices when this month's decision contradicts last quarter's.

Extraction plus retrieval is common; contradiction detection is not, and it is the feature that earns the LLM spend — the graph tells you two of your projects settled the same question differently. Two operational notes worth internalising: the watcher only picks up new changes, so backfill with `cortex ingest` first, and do not run `watch` and `serve` at once because they compete for file events.

What it is

GZOO Cortex watches the directories you register, uses an LLM to pull entities — decisions, patterns, components, dependencies, constraints, action items — out of md, ts, js, py, json and yaml files, infers relationships between them, and lets you query the result in natural language with source citations. It ships an MCP server so Claude Code can query that graph directly.

What you get

12 MCP tools: `cortex_ask` for natural-language questions about your projects, `query_cortex` for structured graph queries, `search_entities` with filters, `find_entity` by name, `session_brief` for a context summary at the start of a session
Contradiction handling — `get_contradictions` lists conflicts the extractor found between decisions, `resolve_contradiction` settles one
Project and ingestion control — `list_projects`, `add_project`, `remove_project`, `ingest_file`, `get_status`
Search blends full-text with vector similarity when embeddings are enabled — configured as a separate provider from your chat model, so chat can run on one vendor and embeddings on another
A web dashboard at `http://localhost:3710` with a D3 knowledge graph, live feed, query explorer and contradiction resolver
Privacy is enforced at the file level: directories marked restricted are never sent to cloud LLMs, `.env`, `.pem` and `.key` files are auto-blocked, and secrets are redacted before any cloud call

Requirements

Node.js 20+ and an LLM API key for cloud modes — Anthropic, Google Gemini, DeepSeek, Groq or any OpenAI-compatible provider. Ollama is required only for `hybrid`, `local-first` or `local-only` routing. Install with `npm install -g @gzoo/cortex`, then `cortex init` walks through provider, key, routing mode, watch directories and a monthly budget cap; `cortex doctor` validates the result. Register projects with `cortex projects add`, backfill existing files with `cortex ingest`, then `cortex serve`. Wire up Claude Code with `claude mcp add cortex --scope user -- npx @gzoo/cortex mcp`. Config lives in `~/.cortex/cortex.config.json`, keys in `~/.cortex/.env`, all data local under `~/.cortex/`. Current release 0.8.1. MIT.

Setup effort

One command plus a key — npm install -g @gzoo/cortex, then supply credentials