cortex-ia
AI Agent Ecosystem Configurator
One command. Any agent. Full SDD stack.
cortex-ia detects your installed AI coding agents and configures them with a complete development ecosystem: persistent memory, SDD workflow, inter-agent messaging, multi-CLI orchestration, and live documentation — all via a single Go binary with an interactive TUI.
Quick Start
# Install via Go
go install github.com/lleontor705/cortex-ia/cmd/cortex-ia@latest
# Install via Homebrew
brew install lleontor705/tap/cortex-ia
# Install via script (Linux/macOS)
curl -sSL https://raw.githubusercontent.com/lleontor705/cortex-ia/main/scripts/install.sh | bash
# Interactive TUI installer
cortex-ia
# CLI: auto-detect agents, full preset
cortex-ia install
# CLI: specific agent + minimal preset
cortex-ia install --agent claude-code --preset minimal
# Preview without changes
cortex-ia install --dry-run
# Show detected agents
cortex-ia detect
cortex-ia injects 5 MCP servers + 19 SDD skills + orchestrator prompts into any supported agent:
| Component |
MCP Tools |
What It Does |
| Cortex |
19 |
Persistent cross-session memory with knowledge graph, FTS5 search, importance scoring |
| ForgeSpec |
15 |
SDD contract validation (Zod), task board with dependencies, file reservation |
| Agent Mailbox |
9 |
Inter-agent messaging, threads, broadcast, request/reply, deduplication |
| CLI Orchestrator |
4 |
Multi-CLI routing (Claude/Gemini/Codex) with circuit breaker, retry, fallback |
| Context7 |
2 |
Live framework and library documentation via MCP |
Plus 3 content components:
| Component |
What It Does |
| SDD Workflow |
9-phase Spec-Driven Development with orchestrator + 19 specialized skills |
| Conventions |
Shared cortex memory protocol + naming conventions for all agents |
| Extra Skills |
Non-SDD utility skills (injected separately from SDD) |
Total: 49 MCP tools, 91% referenced across skills (4 admin-only tools correctly excluded).
Supported Agents
| Agent |
MCP Config |
Prompt Strategy |
Task Delegation |
Sub-Agents |
Slash Commands |
| Claude Code |
Separate JSON files |
Markdown sections |
✅ |
— |
— |
| OpenCode |
Merge into settings |
File replace |
✅ |
✅ |
✅ |
| Gemini CLI |
Merge into settings |
File replace |
— |
— |
— |
| Cursor |
MCP config file |
File replace (.mdc) |
— |
✅ |
— |
| VS Code Copilot |
MCP config file |
File replace |
✅ |
— |
— |
| Codex |
TOML file |
File replace |
— |
— |
— |
| Windsurf |
MCP config file |
Append to file |
— |
— |
— |
| Antigravity |
MCP config file |
Append to file |
— |
— |
— |
Agents with Task Delegation get a multi-agent orchestrator that delegates work to sub-agents. Others get a single-agent prompt that executes SDD phases sequentially.
Presets
| Preset |
Components |
Use Case |
| full |
All 8 components |
Complete ecosystem (default) |
| minimal |
Cortex + ForgeSpec + Context7 + SDD |
Essential SDD workflow (auto-pulls Mailbox via deps) |
| custom |
User-selected via TUI |
Pick exactly what you need |
SDD Pipeline
Spec-Driven Development structures substantial changes through 9 phases:
init → explore → propose → spec → design → tasks → apply → verify → archive
proposal → spec ──┐
↘ ├→ tasks → apply → verify → archive
design ──┘
19 Specialized Skills
| Phase |
Skill |
Role |
| init |
bootstrap |
Detect stack, bootstrap persistence, build skill registry |
| explore |
investigate |
Read codebase, compare approaches, rate effort/risk |
| propose |
draft-proposal |
Create change proposal with scope, risks, rollback plan |
| spec |
write-specs |
Write delta specs with Given/When/Then scenarios |
| design |
architect |
Technical design with architecture decisions |
| tasks |
decompose |
Break specs + design into dependency-ordered tasks |
| apply |
team-lead |
Coordinate parallel @implement agents |
| apply |
implement |
Write production code satisfying specs |
| verify |
validate |
Run tests, generate spec compliance matrix |
| archive |
finalize |
Merge specs, close change, generate retrospective |
Utility Skills: debug, ideate, debate, monitor, execute-plan, open-pr, file-issue, parallel-dispatch, scan-registry
Modern Prompting Techniques
Skills incorporate research-backed techniques for better AI performance:
| Technique |
Applied To |
Impact |
| Chain-of-Verification |
validate |
30-50% fewer hallucinations in verification |
| Constitutional Self-Critique |
implement |
Code critiqued against specs before submission |
| Skeleton-of-Thought |
draft-proposal, write-specs |
Outline → validate → expand reduces omissions |
| Extended Thinking |
architect, decompose |
Explicit trade-off analysis, 2+ alternatives |
| ReAct (Thought/Action/Observation) |
debug |
Grounded debugging with evidence loops |
| Step-Back Prompting |
architect |
Abstract principles before specific design |
| Inline WHY |
orchestrator, all rules |
Motivation on every rule improves compliance |
How It Works
Installation Flow
cortex-ia install
│
├─ Detect system (OS, arch, package manager)
├─ Detect installed agents (scan PATH + config dirs)
├─ Select preset (full / minimal / custom)
├─ Resolve dependencies (SDD → cortex + forgespec + mailbox)
├─ Create backup snapshot (~/.cortex-ia/backups/)
├─ Apply per agent:
│ ├─ Inject MCP configs (strategy-specific: JSON / merge / TOML)
│ ├─ Inject orchestrator prompt (markdown sections / file replace / append)
│ ├─ Write SDD skill files to skills directory
│ ├─ Write shared conventions to _shared/
│ ├─ Write slash commands (OpenCode only)
│ └─ Write sub-agent definitions (OpenCode, Cursor)
└─ Save state (~/.cortex-ia/state.json)
Key Design Principles
- Non-destructive: Uses
<!-- cortex-ia:ID --> markers. Content outside markers is never touched.
- Backup-first: Automatic snapshot before every install with restore capability.
- Idempotent: Running install twice produces identical results with zero file changes.
- Adapter pattern: Each agent implements an interface. Adding a new agent requires zero changes to components.
- Strategy dispatch: MCP injection is template-based — adding a new MCP server is one file.
Documentation
| Doc |
Description |
| Installation |
All installation methods, prerequisites, platform notes |
| Agents |
Per-agent configuration details, paths, strategies |
| Components |
Component catalog, dependencies, what each injects |
| SDD Workflow |
9-phase pipeline, commands, contract validation, prompting techniques |
| Architecture |
Codebase structure, patterns, testing, contributing |
| Configuration |
Presets, CLI flags, state management, backup/restore |
Prerequisites
- Go 1.22+ — for building cortex-ia
- Node.js 18+ with
npx — for npm-based MCP servers (forgespec, mailbox, orchestrator, context7)
- Cortex binary —
go install github.com/lleontor705/cortex/cmd/cortex@latest or brew install lleontor705/tap/cortex
- At least one supported agent installed
License
MIT