Open Source · Built with AI Agents

One person.
Infinite agents.

Cortex is a personal agent infrastructure stack. Four integrated systems that let a single person command, orchestrate, create tools, and capture wisdom — all through AI agents.

Get Started See Architecture
cortex — zsh
# Message your AI from your phone
telegram> refactor the auth module, add proper error handling
Dispatcher: on it (myapp) — 3 agents mobilized

# Agents discover and delegate to each other
a2a-hub> code-reviewer registered · test-runner registered
a2a-hub> task delegated: review → code-reviewer

# Need a new tool? Forge builds it on the fly
forge> create "validate JWT tokens with role-based access"
Generated · 12 tests passed · Installed to MCP

# Every session becomes structured wisdom
vibe-replay> Session captured: 4 decisions · 2 learnings · replay ready
The Paradox

It's 2026. AI can code, research, analyze, and deploy.
So why does it still feel like juggling?

Claude Code is extraordinary. MCP servers are powerful. Skills and plugins keep multiplying. Each tool, individually, is the best it's ever been.

But string them together across a real workday, and something breaks down.

A Real Day

Sound familiar?

9:15
Open laptop. Start Claude Code. "Refactor the data pipeline."
Works beautifully. AI reads 20 files, plans the refactor.
9:47
Need a custom CSV parser. Claude can't find one from last week's session.
Write it from scratch. Again. CONTEXT LOST
10:22
Walking to get coffee. Great idea for the API design.
Can't execute from phone. Text yourself a note. UNREACHABLE
11:08
Back at laptop. "Where was I?" Re-read the codebase.
20 minutes to rebuild the context you already had. START OVER
2:30
Refactoring done. 24 tests passing. Solid work.
But what decisions were made? Why 3 modules instead of 2? Already gone. WISDOM LOST
3:15
Colleague: "How did you approach that refactor? I need to do something similar."
You can't reconstruct the reasoning. It lived and died in one session. NOT SHAREABLE
Every day. Every session. Starting from zero.
The Diagnosis

The tools are there. The glue isn't.

The Brain
LLMs are superhuman. Claude, GPT, Gemini — reasoning, coding, analysis. Solved.
The Hands
MCP servers, skills, plugins, tool use. Agents can read, write, search, deploy. Solved.
?
The Nervous System
What connects them? What makes them work as one organism instead of scattered limbs?
We have superhuman intelligence and powerful tools.
What's missing is the infrastructure between them — the layer that lets you command from anywhere, agents discover and delegate to each other, new tools emerge on demand, and every session compounds into wisdom.
The Shift

Same person. Same tools. Different infrastructure.

A Thursday afternoon — Without Cortex
Y
Opens Claude Code on laptop...
"Refactor the data pipeline, add error handling"
C
I'll read the codebase first...
[reads 20 files, plans refactor, starts working]
— 32 min later —
C
I need a JWT validation helper for the auth middleware.
Y
I built one last week in a different session... let me re-describe it.
— 15 min rebuilding what existed —
!
Phone buzzes. Walking to a meeting. Great idea for the API design. Can't do anything about it.
— 45 min meeting —
Y
Back at laptop. "Where was I?" Session context is stale. Re-reading everything.
— 2 hours later, task done —
!
Colleague asks how you approached it. You can't remember the key decisions. The reasoning is gone.
A Thursday afternoon — With Cortex
Y
Sends Telegram message from phone while walking...
"Refactor the data pipeline, add proper error handling"
D
On it (myapp). Detected project, spawning session.
— Agents working —
H
Agent needs JWT validator. Found Forge via A2A Hub. Delegating.
F
Built jwt_validator. 12 tests passing. Installed to MCP.
— You're still in your meeting —
Y
Quick Telegram reply between meetings...
"Also add role-based access control to that validator"
D
Got it. Resuming session. Adding RBAC.
— 20 min later —
D
Done. Refactored into 3 modules, 24 tests, new JWT tool installed. Pushed to git.
V
Session captured: 4 decisions, 2 learnings, 1 new tool. Replay ready. Shareable link generated.
Y
Forwards the replay link to colleague. "Here's exactly how and why."

Same person. Same AI. Same tools.
The difference is the layer between them — the infrastructure that turns isolated capabilities into a unified, compounding system.

Architecture

Four systems. One brain.

Each component handles a distinct cognitive function. Together, they form a complete agent operating system — from receiving commands to building tools to remembering what worked.

YOU
Phone / Desktop / Anywhere
Telegram / CLI / Natural Language
DISPATCHER
Command Layer
Classify → Route → Execute → Reply
Mobile JARVIS Always-on Daemon Zero Dependencies
A2A HUB
Communication Layer
Agent Registry · Task Routing · Discovery
WebSocket Protocol MCP Bridge
FORGE
Creation Layer
Describe → Clarify → Generate → Test → Install
CLI + MCP Server Auto Iterate
VIBE REPLAY
Memory Layer
Capture Events → Analyze → Extract Insights → Share
Claude Code Hooks HTML Timeline Wisdom Aggregation
Components

Built to work alone. Better together.

Each component is a standalone, open-source project. Use one, use all four. They're designed to compose naturally.

D

Dispatcher

Command Layer

Your mobile JARVIS. Message Claude from your phone via Telegram. It classifies, routes, and manages concurrent tasks — always on, always ready.

H

A2A Hub

Communication Layer

Agent-to-Agent protocol hub. Agents discover each other, delegate tasks, and collaborate through a lightweight WebSocket-based protocol.

F

Forge

Creation Layer

Self-evolving tool agent. Describe what you need in plain English — Forge builds it, tests it, and installs it into your toolkit.

V

Vibe Replay

Memory Layer

Not just logs — structured wisdom. Captures coding sessions, extracts decisions and learnings, generates beautiful shareable replays.

How It Works

From thought to execution to wisdom

A real workflow: you're on your phone, you have an idea, and your agent infrastructure handles the rest.

01

You send a message from your phone

"Refactor the data pipeline, add proper error handling and tests."
Sent via Telegram. You're on the subway. No laptop needed.

02

Dispatcher classifies and routes

Detects project myapp, sets working directory, classifies as engineering task. Spawns a Claude Code session with 50 max turns and your memory context. Sends you: "On it (myapp)."

03

Agent discovers it needs a CSV parser tool

During refactoring, Claude realizes it needs a JWT validation tool. Through A2A Hub, it finds Forge and delegates: "Build a JWT validator with role-based access control."

04

Forge builds, tests, and installs the tool

Forge generates the validator, creates 12 tests, runs them, all pass. Installs as an MCP tool. The refactoring agent now uses it directly.

05

Task completes. Results sent to your phone

"Done. Refactored into 3 modules, added error handling and 24 tests. Also created a new jwt_validator tool (installed to MCP)."

06

Vibe Replay captures everything

The entire session is captured: 4 key decisions (why 3 modules instead of 2), 2 learnings (JWT token edge cases), 1 new tool created. HTML replay generated. Wisdom aggregated.

Use Cases

What one person can do now

Cortex isn't theoretical. It's built by a researcher who uses it daily to run experiments, ship projects, and automate everything in between.

One-Person Research Lab

Run literature reviews, design experiments, analyze data, and draft papers — all orchestrated by agents. Your phone becomes your lab terminal.

Dispatcher A2A Hub Vibe Replay

Rapid Tool Building

Hit a wall during development? Describe the tool you wish existed. Forge builds it, tests it, and injects it into your workflow in minutes.

Forge A2A Hub

Vibe Coding Knowledge Base

Every AI coding session generates structured learnings. Over time, you build a personal knowledge base of what works — shareable with your team or community.

Vibe Replay

Mobile-First Engineering

Commuting? At dinner? Send a Telegram message and your agents handle the rest. Concurrent tasks, progress updates, smart follow-ups — like texting a senior engineer.

Dispatcher

Multi-Agent Workflows

Chain specialized agents together. One researches, another codes, a third reviews, and a fourth writes docs. All coordinated through a shared protocol.

A2A Hub Forge Dispatcher

Personal Automation

Flight lookups, visa research, calendar management, daily book generation — set up once, runs forever as an always-on daemon.

Dispatcher Forge
Quick Start

One install. Everything works.

Cortex is one package. Install it, run the setup, and all four systems are ready.

1 — Install Python 3.11+
pip install cortex-agent
2 — Setup (interactive, one time)
cortex init

# Walks you through:
> Telegram bot token: sk-...
> Your chat ID: 1234567
> Project directories to register: ~/projects/myapp, ~/work/api
> Enable Vibe Replay hooks? Y

✓ Config saved to ~/.config/cortex/config.yaml
✓ Claude Code hooks installed
✓ Ready to go.
3 — Start
cortex start

✓ Dispatcher running — send a message on Telegram
✓ A2A Hub listening on :8765
✓ Forge MCP server registered
✓ Vibe Replay capturing sessions

# Or install as always-on daemon (starts on boot)
cortex install # macOS LaunchAgent / Linux systemd
4 — Use
# From your phone (Telegram)
You: refactor the auth module, add error handling
Cortex: On it (myapp). Will report back when done.

# Need a new tool?
cortex forge "validate JWT tokens with role-based access"
✓ Generated · 12 tests passed · Installed to MCP

# Review what happened
cortex replay latest
✓ Opening interactive replay — 4 decisions, 2 learnings

# See what you've learned over time
cortex wisdom