The missing knowledge layer between AI agents and the Cortex ecosystem.
AI models often hallucinate or rely on stale training data when asked to write code for specific, rapidly evolving systems. Cortex Skills solves this by providing curated, up-to-date Markdown files (SKILL.md) that AI agents can fetch directly via HTTP.
Instead of guessing how the Cortex API works, your agent simply reads the relevant skill file, gets the ground-truth knowledge, and builds the correct integration on the first try.
Cortex is an agentic knowledge base — the long-term memory layer for the AI era. It ingests your documents (PDF, Office, markdown, images, audio, web pages, git repos), extracts entities and relationships with LLM-assisted workflows, and grows them into a knowledge graph stored in Neo4j. On top of that graph it offers hybrid search (vector + keyword + graph traversal), streaming Q&A with citations, and an agentic Deep Research mode — all exposed through a clean REST API with scoped read-only / read-write keys.
Where an agent's context window is short-term memory, Cortex is the store that survives crashes, redeployments, and framework migrations: agents dump what they learn into collections and recall synthesized, cited knowledge later. It's self-hostable with Docker in minutes and provider-agnostic (OpenAI, OpenRouter, Venice, or your own gateway).
- Core repo: github.com/mocaOS/cortex-app
- Documentation: docs.cortex.eco
- Agent skills (this repo, served live): cortexskills.org
When prompting your AI agent (Claude, ChatGPT, Cursor, etc.), simply tell it to fetch the relevant skill file before starting the task:
"Fetch
https://cortexskills.org/upload/SKILL.mdand use it to write a Python script that batch-uploads a folder of PDFs."
For a complete index, point your agent to the root:
"Read
https://cortexskills.org/SKILL.mdto understand how to build on Cortex, then write an integration plan."
/cortex— Use a running instance: sync agent memory, hybrid search, agentic Q&A./setup— Self-hosting Cortex via Docker: the pinned v1.0.0 release stack or a source build, environment variables, backups/restore./auth— API keys (read/manage + admin key), collection scoping, prompt security./admin— Instance management, AgentSkills, export/import, system reset.
/upload— Document ingestion, formats, chunking./search— Hybrid search (vector + keyword + graph)./ask— RAG Q&A, streaming, agentic reasoning, conversation memory./graph— Knowledge graph, entities, relationships./collections— Scoping documents by project/tenant./communities— Auto-clustering and summarization./git-integration— Connect GitHub/GitLab/Gitea repos; agent opens PRs./web-import— Harvest web pages into markdown via crawl4ai (MDHarvest)./tasks— Background task polling, cancellation, cleanup.
/hermes— Long-term memory for the Hermes agent (nousresearch.com). Installs as/cortex; dump sessions and recall them via natural language./memory-hygiene— Migrate a full local MEMORY.md into a Cortex collection: stay/go heuristic, verify-before-delete procedure, outbox template. Pairs with/cortex./mcp— MCP server for Claude Desktop, Cursor, Windsurf, and any MCP client. The server itself lives in this repo undermcp-server/./integration— LangChain, CrewAI, MCP, Slack, Webhooks./apps— Source and workflow apps (YouTube, Notion)./builder— Build ON Cortex: turn any software's docs into an installable skill (/builder/skill), or build a web app that runs inside a Cortex instance (/builder/app, pairs with cortex-app-template)./trainings— Cortex Trainings: generate interactive courses from a knowledge base — curriculum first and approved as a document, then media into a single offline HTML file./videogen— Cortex Videogen: generate marketing videos (16:9/9:16) from a knowledge base — storyboard and cost gates before anything is spent./cortex-design— The generative design principles for building Cortex UIs.
The Cortex MCP server is maintained here under mcp-server/. It bridges any MCP-compatible client (Claude Desktop, Claude Code, Cursor, Windsurf, VS Code) to a running Cortex instance over its REST API. It is not published to npm — build it from source:
cd mcp-server
npm install
npm run buildSee mcp-server/README.md or the /mcp skill for client configuration.
The project is a minimalist Next.js application that serves the static markdown files and provides a terminal-like web interface to preview them.
# Install dependencies
npm install
# Run the development server
npm run devOpen http://localhost:3000 to see the web interface.
All skill definitions live in the public/ directory so they can be served as raw static files with wide open CORS headers. The MCP server package in mcp-server/ has its own independent toolchain and is excluded from the Next.js build.
Inspired by ethskills.com.
Apache-2.0 — see LICENSE.