CORTEX v1.1.0

First public release.

What CORTEX is

An Obsidian-style Markdown vault your AI agents read and write over MCP — plus an orchestrator that
splits coding work by cost: a free or on-device model implements, your expensive model only reviews
the diff.

Highlights

Vault

• Markdown notes with [[wiki links]], FTS5 search and a backlink graph
• A write-only .md mirror you can open in Obsidian
• Semantic search powered by an on-device embedding model

Shared memory for agents

• ~20 cortex_* MCP tools plus a REST API on localhost:7777
• Standing rules from the PROTOCOL tab are injected into every agent's instructions on connect, so
  you stop re-pasting your stack and conventions into every session
• Every note write and task move is attributed to the agent that did it

Orchestrator

• planner → coder → security → QA → planner, with manual gates wherever you want to approve by hand
• Per-role agent and model assignment, plus a per-task override when you move a card
• Coding agents run in an isolated sandbox; the launcher computes the diff, runs the type-check gate,
  and only then applies the result to your repo
• Per-role spend caps enforced at launch

On-device AI

• GGUF weights in-process via node-llama-cpp — no Ollama, no server, no network port
• Turns a rough sentence into a task spec with acceptance criteria
• Deterministic Librarian scanners for exposed secrets, stalled work, contradictions and
  prompt-injection attempts

Install

macOS on Apple Silicon, Node.js 20+.

git clone https://github.com/blackwolffdev/cortex.git
cd cortex && npm install && npm run app

Known limits

• macOS/arm64 only — Linux and Windows is #2
• English UI only — i18n is #1
• No test suite yet — #3
• opencode runs without an MCP config — #5

Contributors welcome. Every one of those is a well-scoped starting point.