Cortex

One always-on brain, built on Amplifier

Built with Amplifier

Cortex is one always-on brain that decides which notifications actually deserve you

A hub-and-spoke personal-AI platform: one persistent server hosting pluggable Domain Experts. Its first expert already runs.

And it started as one person's idea — with the evidence to prove it.

Cortex Core

Always-on hub — HTTP/WebSocket API + web UI on port 19420

Attention Firewall

First Domain Expert — MVP working, scoring live notifications

The evidence

This started as one person's idea — and it already works

Cortex is a solo project by Brian Krabach at Microsoft: 29 commits, 100% his. Its first expert is already an MVP scoring live notifications.

Attention Firewall · MVP working

So what makes a filter like this necessary in the first place?

29

commits, 100% by Brian Krabach

1

sole author of bkrabach/cortex

MVP

notifications flowing, LLM scoring live

The problem

Today's assistants are stateless and siloed — so every ping shouts equally

They carry no memory across interactions and treat every notification as equally urgent. The result is noise: you drown in pings, and nothing judges what actually deserves you.

Cortex's answer is an expert whose only job is that judgment.

The first Domain Expert

The Attention Firewall captures your notifications and lets an LLM judge each one

Windows notifications are captured, ingested to the server, and stored. An Anthropic model scores each one for relevance using recent conversation context.

Then it does something a plain inbox never does — it decides.

1Capture — Windows notifications via pywinrt
2Store — server ingestion + SQLite storage
3Score — LLM relevance 0–1 with 24h / 10-msg context

How it decides

One score, 0 to 1, decides whether a notification reaches you

Instead of equal-urgency noise, every notification gets a single relevance score — and that score routes it to one of three outcomes.

This judgment is only possible because of what Cortex is built on.

< 0.3Suppress — you never see it
0.3–0.6Summarize — folded into a digest
≥ 0.6Push — WebSocket devices + ntfy.sh mobile

The foundation

Cortex is built on Amplifier primitives — that's why it extends

A long-running Amplifier session is the core; each Domain Expert is a bundle plus a session. The cortex-a2a bundle runs on loop-streaming, context-simple, and claude-sonnet-4-5.

That foundation lets the Firewall do more than notify.

# bundles/cortex-a2a.md
orchestrator: loop-streaming
context: context-simple
model: claude-sonnet-4-5
hooks-a2a-server:
  port: 8214
  agent_name: "cortex"

The payoff

At 0.9, Cortex stops notifying you and acts on your behalf

When a notification scores ≥ 0.9 urgency, Cortex fires an autonomous A2A message to peer agent ai-os — a fire-and-forget task across a 5-agent mesh, no waiting for you.

One vision, six months old, is now doing this for real.

0.9

≥ 0.9Broadcast — autonomous A2A message to peer agent ai-os
mesh5 agents: ai-os, lifeline, lifeline-demo, hive-slack, cortex (port 8214)

What you keep

A six-month-old vision is now a working hub-and-spoke platform

The Cortex vision predates Amplifier by ~6 months (dated 2025-03-05). Today it's a real system — and the pattern scales: more Domain Experts are planned next.

One brain. Pluggable experts. The Attention Firewall was just the first.

✓Attention Firewall — MVP working today
→Email Manager — planned (future)
→Calendar — planned, with a 5-level autonomy progression Observer → Autonomous

Sources

Research Methodology

Data as of: February 26, 2026 (latest commits in bkrabach/cortex).

Feature status: Attention Firewall — MVP working; A2A autonomy — implemented & test-backed; Cortex Core & future experts — in progress / planned.

Primary source: local monorepo bkrabach/cortex (README, ROADMAP, CONTEXT_TRANSFER.md, vision docs, bundles, tests). Claims re-derived from git/grep/wc output.

Commands run:

git -C bkrabach/cortex shortlog -sne HEAD · git rev-list --count HEAD — 29 commits, 100% Brian Krabach
grep -n 'Decision:' bkrabach/cortex/CONTEXT_TRANSFER.md — push ≥0.6 / summarize 0.3–0.6 / suppress <0.3; ≥0.9 triggers A2A
grep -n '19420\|8214' CONTEXT_TRANSFER.md · git show cd38e66 --stat — ports 19420 (API/UI) & 8214 (A2A)
sed -n '1,20p' context/a2a-network.md — 5-agent mesh (ai-os, lifeline, lifeline-demo, hive-slack, cortex)
sed -n '1,45p' bundles/cortex-a2a.md — loop-streaming, context-simple, claude-sonnet-4-5
wc -l bkrabach/cortex/tests/*.py — 572 test lines · grep 'predates' ROADMAP.md — vision predates Amplifier ~6 months (2025-03-05)

Gaps: Cortex submodules (amplifier-app-server, attention-firewall client, macOS client) have empty local working trees; filenames (pywinrt, llm_scorer.py, ntfy.sh push) are quoted from CONTEXT_TRANSFER.md as documented status, not runtime-verified. The primary-source repo kenotron-ms/amplifier-presentation exists on GitHub but was not cloned locally.

Primary contributor: Brian Krabach (brkrabac@microsoft.com) — 29 commits, 100%. Amplifier co-authored the deck, bundle, and branding commits.