What happens when you put an AI agent framework
inside a single HTML file
Sign up, verify email, set up a profile
Generate, store, rotate, protect
Always-on connection to a remote server
Monthly fees, usage limits, billing anxiety
That your data stays private on someone else's server
What if none of that was necessary?
Open it in Chrome. No server. No API key. No internet after first model download.
The AmplifierBrowser facade boots Python inside WebAssembly via Pyodide. The same amplifier-core kernel that powers terminal sessions initializes in your browser tab.
WebLLM loads a quantized model directly onto your GPU through WebGPU. No network calls. No intermediary. Your prompt goes straight to local silicon.
Microsoft's compact powerhouse. Strong reasoning for its size, ideal for general Q&A, tutoring, and code assistance.
Meta's speed-optimized model. Quick responses for conversational use, summaries, and lightweight tasks.
Alibaba's multilingual model. Higher quality output with support for many languages. Needs more VRAM but delivers more capable results.
A teacher sends students an HTML file. Inside is a specialized tutor that knows the curriculum, answers questions, and works offline in any classroom.
A company distributes a portable knowledge base. No server infrastructure. No ongoing costs. Every employee gets AI on their laptop.
A journalist in a sensitive region has AI assistance without any network traffic. No server logs. No API calls to intercept. Complete operational security.
A developer prototypes agent behavior in a browser tab. No deployment pipeline. No cloud account. Open a file, iterate, ship.
macOS, Linux, Windows — the classic CLI experience
amplifier-coreContainerized workloads with full isolation
amplifier-coreRemote machines and build servers
amplifier-coreWebAssembly — the new frontier
amplifier-coreSame protocols. Same composition model. Same bundle format. The platform() function in the kernel has 'wasm' as a first-class return value — the browser was a design target from day one, not an afterthought.
The webruntime-developer agent doesn't just build browser AI apps — it tests them autonomously with Playwright and WebGPU flags. No untested code reaches the user.
Playwright launches Chromium with WebGPU flags enabled. The agent generates tests, runs them headless, reads console output on failure, and iterates until everything works.
When a test fails, the agent reads the console logs, diagnoses the issue, fixes the code, and re-runs. This loop continues until the app works correctly — no human debugging required.
Attach an AI assistant to a message. The recipient opens a file and has a working AI.
Carry an AI in your pocket. Works anywhere with a modern browser and a GPU.
Host AI on a $0/month static site. No backend. No database. No ops team.
Drop an AI into documentation, a wiki, an internal portal — it just works.
No accounts. No infrastructure. No ongoing costs.
The file IS the application.
Data as of: February 26, 2026
Feature status: Active — code present in production repositories with active development
Research performed:
amplifier_webruntime.py — confirmed imports from amplifier_core.interfaces (Provider, ContextManager, Tool, Orchestrator)amplifier_env_common/protocol.py — confirmed 'wasm' as first-class platform() return valuewebruntime-developer in amplifier-foundation confirmed with Playwright + WebGPU testing workflowRepositories: amplifier-core (kernel), amplifier-foundation (webruntime-developer agent), webllm bundle (browser LLM provider)
Model VRAM requirements: From WebLLM model catalog documentation. Actual VRAM usage may vary by quantization and browser.
Primary contributors: Not determined — research did not extract individual contributor attribution from git history for this deck. The browser runtime spans multiple repos and contributors.
Gaps: No independent VRAM benchmarks were run. Model parameter counts are from published model cards. Exact initialization line count (~6 lines) is from the AmplifierBrowser facade API surface, not a formal LOC metric. Individual contributor commit shares were not analyzed.
Every other AI product requires infrastructure.
Amplifier's browser runtime requires a browser.
That's not a limitation — it's liberation.
Explore Amplifier on GitHub →