The Server That
Runs Amplifier

10K lines of always-on AI runtime

The Legend

A story spread that the code never confirms

The pitch: four casual prompts, no code written by hand, and out came a serverless, browser-native app running a whole AI agent framework on your machine.

Seductive. But does the repository actually hold that? Let's open it.

The Artifact

The repo behind the story is real — and correctly identified

The git remote confirms it: this is bkrabach/amplifier-app-server, authored and solely committed by Brian Krabach. We are testing the actual code, not a straw man.

With the real artifact in hand, we can hold it up against the legend.

$ cd amplifier-app-server && git remote -v
origin  git@github.com:bkrabach/amplifier-app-server.git (fetch)
origin  git@github.com:bkrabach/amplifier-app-server.git (push)

The Contradiction

The code describes itself as the opposite of the pitch

Its own description reads: "Always-on Amplifier server runtime — hosts persistent AI sessions with HTTP/WebSocket API." Not serverless. Not browser-native-local. A server.

So what, concretely, is this thing that the story mislabeled?

What It Really Is

"Cortex Hub": a substantial, real server project

The internal name is Cortex Hub — an attention and notification management platform. It's not a toy: it's built on FastAPI, uvicorn, and WebSockets.

Scale that size settles the question of "serverless" for good.

10,240

lines of Python

39

Python files

68

HTTP/WebSocket routes

14

server-hosted UI files

Not Serverless

It's deployed as a long-running system service

A systemd unit runs amplifier-server on port 19420 with Restart=on-failure. An always-on service is the definition of not serverless.

But here's the twist that makes the truth better than the legend.

# service/cortex-server.service
[Service]
Type=simple
ExecStart=.../amplifier-server run --port 19420
Restart=on-failure

The Twist

This server doesn't just host chat — it runs Amplifier itself

It imports amplifier_foundation, calls load_bundle, and runs real sessions via create_session and session.execute — pulling microsoft/amplifier-core and microsoft/amplifier-foundation.

That reframes everything: the corrected story outshines the myth.

# session_manager.py
from amplifier_foundation import Bundle
from amplifier_foundation import load_bundle
await prepared.create_session(...)
result = await session.execute(prompt)

The Reveal

The unglamorous truth beats the legend

Not a four-prompt serverless toy — a real 10,240-line always-on runtime, on port 19420, that loads bundles and runs Amplifier sessions itself. Built by describing intent to Amplifier.

The evidenced achievement is bigger than the myth it replaces.

What You Keep

Check the repo before the legend

The "four prompts / no code / single session" origin has zero supporting evidence in the repo — and the honest story is the better one.

Verify claims, and let the real artifact carry the story.

0 matches for "serverless" in the repo
0 matches for "browser-native"
0 matches for "four prompt"
1 commit in the clone — a bug fix, not an origin

Sources

Research Methodology

Source repo: bkrabach/amplifier-app-server (git remote confirmed)

Data as of: only git commit dated 2026-02-26; earliest in-content date 2026-01-16

Feature status: Working, actively developed server project

Commands run:

Remote: git remote -v
Description: cat pyproject.toml and grep README.md
Service: cat service/cortex-server.service
Amplifier usage: grep -nE 'from amplifier|Bundle|create_session|execute' src/amplifier_server/session_manager.py
Size: find src -name '*.py' | wc -l (39); find src -name '*.py' -exec cat {} + | wc -l (10,240)
Routes: grep -rhoE '@(app|router)\.(get|post|delete|put|websocket)\("[^"]*"' src/amplifier_server/api/ | sort -u | wc -l (68 unique / 71 total)
Static UI: find static -type f | wc -l (14); grep '<title>' static/index.html
History: git rev-list --all --count (1); git log --format='%h %ai %an <%ae> %s'
Negative checks: grep -rai 'serverless', 'browser-native', 'four prompt' (0 matches each)

Gaps: Exact project creation date unknown (file mtimes reflect the 2026-07-08 clone date). The "four prompts / no hand-written code / single session / serverless / browser-native" origin narrative has no supporting evidence and is not stated as fact.

Primary contributor: Brian Krabach <brkrabac@microsoft.com> — author / sole committer.

The Server ThatRuns Amplifier