Cortex
GitHub
Model registry refreshed daily

Write your AI rules once.
Every coding agent follows them.

Cortex compiles one .cortex/ folder into the native instruction files of Claude Code, Codex, Cursor, Copilot, Gemini CLI, Windsurf, Kiro and Antigravity — formatted for each model, fitted to each tool’s limits, and checked for drift in CI.

Try it in the browser Plan it in the Lab first
$npm install -g github:Phani3108/Cortex
Zero dependencies Runs locally MIT licensed Never overwrites your files
.cortex/rules/project.mdsource
## Architecture
- Business logic lives in src/services/
- ! Never commit secrets or .env files

## Testing
- Every bug fix ships with a failing test first
──cortex compile──▶
CLAUDE.mdClaude Code
AGENTS.mdCodex +20
.cursor/rules/cortex.mdcCursor
copilot-instructions.mdCopilot
GEMINI.mdGemini CLI
.windsurf/rules/cortex.mdWindsurf
.kiro/steering/cortex.mdKiro
.agents/skills/*/SKILL.mdAgent Skills
Why Cortex

Your team uses four AI tools. They are following four different rulebooks.

Every tool reads a different file, in a different format, with different limits — and the copies drift the week after you write them.

Without Cortex
  • The same rules pasted into CLAUDE.md, .cursor/rules, Copilot instructions and AGENTS.md
  • Windsurf silently cuts rule files past 12,000 characters; Antigravity past 24 KB
  • Tools that read AGENTS.md and their own file load your rules twice
  • Nobody notices the copies drifted until an agent does the wrong thing
With Cortex
  • One markdown source in .cortex/, reviewed like code
  • Hard limits enforced by keeping your highest-impact rules, whole — and telling you what didn’t fit
  • Path-scoped rules become native globs; skills become portable SKILL.md
  • cortex compile --check fails CI the moment anything drifts
Playground · runs the real compiler

Edit the source. Watch every tool’s file update.

This is the same engine the CLI ships, running in your browser. Nothing is uploaded. Try adding scope: frontmatter, a ! critical rule, or switching tools on and off.

Source 

            

            

              
              
              
              
            

          

        

      

      
Token counts are estimates (characters ÷ the model family’s average characters-per-token). Limits come from each vendor’s documentation — see the tool matrix.

    

  


  
  

    

      

        
How it works

        
Three commands. Then it runs itself.

      

      

        

          
1

          
Init

          
Detects your stack and the AI tools you already use, and imports hand-written CLAUDE.md / AGENTS.md / Cursor / Copilot files so nothing is lost.

          
cortex init

        

        

          
2

          
Compile

          
Writes each tool’s native files. Refuses to overwrite files a human wrote or edited, and cleans up files it no longer generates.

          
cortex compile

        

        

          
3

          
Guard

          
Commit the sources and the outputs (cloud agents read them from the repo). CI fails when they drift; cortex learn turns your edits into rule suggestions.

          
cortex compile --check

        

      

    

  


  
  

    

      

        
Supported tools

        
Native output for every tool — verified against vendor docs.

        
Paths, frontmatter and limits live in one table in the engine. When a vendor changes its conventions, one edit updates the CLI, the docs and this page.

      

      

        

          
ToolAlways-on rulesPath-scoped rulesSkillsLimitDocs

          
Loading…

        

      

      

    

  


  
  

    

      

        
What’s inside

        
A build system for AI context.

      

      

        

          

          
Model-aware formatting

          
XML sections for Claude, numbered lists for GPT, context-first for Gemini, compact constraints for reasoning models. Set a model per tool and the output follows.

        

        

          

          
Budget-aware, never truncated

          
Hard limits are met by keeping whole, high-impact rules — critical rules (- !) first — and listing exactly what was left out.

        

        

          

          
Path-scoped rules

          
Add scope: ["src/**/*.tsx"] to a rule file. Cortex emits Cursor globs, Copilot applyTo, Claude paths, Kiro fileMatch and Windsurf glob triggers.

        

        

          

          
Agent Skills, portable

          
Skills compile to the open SKILL.md standard in .claude/skills, .agents/skills and .kiro/skills — loaded on demand, not stuffed into every prompt.

        

        

          

          
Safe by default

          
Hand-written or hand-edited files are never overwritten without --force. cortex init imports what you already have first.

        

        

          

          
Live model data

          
Pricing and context windows for 200+ models, refreshed daily by a GitHub Action. cortex cost tells you what your always-on context really costs.

        

      

    

  


  
  

    

      

        

          
Model registry

          
Current models, current prices.

          
Loading…

        

        
        Browse all models →
      

      

    

  


  
  

    

      

        

          
Give every agent the same brain.

          
Install, run cortex init in a repo, and commit the result. Two minutes.

        

        

          

            $npm install -g github:Phani3108/Cortex
            
            
          

          

            $cortex init && cortex compile