Amplifier Stories

AI as a Document

A whole agent in one HTML file

The gap

Every AI product makes you rent someone else's server — what if the agent were just a file you open?

Accounts, API keys, subscriptions, and a round-trip to someone else's machine. All of that is infrastructure standing between a person and a model — not the intelligence itself.

So Amplifier asked: what if the whole thing were a document?

The proof, up front

It already exists: one HTML file that runs an agent in four languages.

The polyglot demo is a single self-contained HTML file bundling an AI agent with tools in Rust, TypeScript, Python, and Go — running entirely in the browser.

A committed, measured artifact — not a mockup.

1

HTML file

4

languages: Rust, TypeScript, Python, Go

4.5 MB

measured on-disk artifact (4,680,357 bytes)

The catch

A self-contained file is worthless if the intelligence still phones home.

One API key, one server round-trip, and the promise collapses — you're back to renting someone else's machine. Self-containment only matters if inference is local too.

Which raises the question: where does the model actually run?

Meet WebAmp

WebAmp installs the real amplifier-core kernel — it doesn't reimplement it.

A sibling proof-of-concept that runs Amplifier entirely in the browser via WebGPU and Pyodide, installing the genuine amplifier-core 1.0.0 wheel with micropip.

So what runs in-browser is the actual kernel — setting up how inference stays local.

1.0.0

real amplifier-core wheel, installed via micropip in Pyodide

80 KB

amplifier_core wheel (80,172 bytes)

The browser shim imports the same interfaces: Provider, ContextManager, HookRegistry, events from amplifier_core.

Where inference runs

The LLM runs on your GPU — no API key anywhere in the client.

Inference happens locally via WebLLM on WebGPU. A grep for api-key, authorization, bearer, openai, and anthropic across the client code returned no matches.

The file owes nothing to a remote service. Now: how do four languages fit in one file?

# client-side model, on your GPU
import 'https://esm.run/@mlc-ai/web-llm'
CreateMLCEngine('Qwen3-4B-q4f16_1-MLC', ...)

# grep api-key|authorization|bearer
# |openai|anthropic  ->  no matches

Four languages, one file

The Go document_builder is compiled to WASM and embedded in the HTML.

document_builder is the first non-Python, non-Rust module in the Amplifier ecosystem. Its WASM build is embedded so all four languages ship inside the one file.

Now open it — and watch everything boot.

1st

polyglot (Go) Amplifier module

3.2 MB

Go WASM binary embedded (3,225,021 bytes)

The payoff

Open the file and five runtimes boot in order — verified.

A single 4.5MB HTML file boots its runtimes and answers on your GPU, backed by 168 test functions and a 49-test end-to-end gate on the assembled HTML.

That turns "toy demo" doubt into proven quality — with one honest caveat.

1Rust WASMagent kernel
2Go WASMdocument_builder
3TypeScripttools
4PyodidePython analyzer
5WebLLMlocal model

The honest edge

First run still needs a local http server and a CDN model download.

"Email one file, double-click, fully offline" is the direction this code is driving toward — the proven trajectory of "AI as a document," not yet a literal file:// double-click.

Every commit here was authored and tested with the Amplifier agent itself.

Today: served over http:// (localhost); first load downloads the model from a CDN.
Aspiration: a file you open, fully offline — the direction the code moves toward.
Built by the agent: all 22 polyglot + all 3 WebAmp commits carry Amplifier trailers.

Sources

Sources & Research Methodology

Data as of: February 19–20, 2026 · Feature status: Working demos / proofs of concept (not shipped products)

Repos: bkrabach/amplifier-polyglot-demo, bkrabach/amplifier-app-webamp, bkrabach/amplifier-module-tool-document-builder-go; kernel: amplifier-core

Commands run:

gh repo view bkrabach/amplifier-polyglot-demo --json description — "Four-language AI agent in a single HTML file"
ls -la html/amplifier-polyglot-agent.html — 4,680,357 bytes (4.5MB)
grep -in 'webllm|Qwen3' bridge.js and grep -in 'api.key|authorization|bearer|openai|anthropic' — no key matches; model Qwen3-4B-q4f16_1-MLC
ls -la public/wheels/ + grep 'micropip' src/main.js — amplifier_core-1.0.0-py3-none-any.whl (80,172 bytes)
ls -la html/go_wasm/document_builder.wasm — 3,225,021 bytes (~3.2MB)
grep -rn 'def test_' html/tests/ | wc -l — 168; commit c8cc5dc — 49 E2E / 134 total tests pass
git log --format='%b' | grep -c 'Amplifier' — all 22 polyglot + all 3 WebAmp commits carry Amplifier trailers

Gaps / honesty: Both repos currently require serving over http:// (localhost), not file://; "fully offline" only works after a first-run CDN model download (~multi-GB). amplifier-core's "~2,600-line kernel" is the README's documented figure, not a re-measured count. Requires Chrome/Edge 113+ (WebGPU).

Primary contributors: Brian Krabach (brkrabac@microsoft.com) — sole human author of all three repos; every commit co-authored/generated by the Amplifier agent.