J.A.R.V.I.S.

A browser voice assistant with an Iron Man holographic interface. Say "Hey Jarvis", he wakes, listens, and does real things through your tools — searches the web, generates images, drives your phone, reads your mail. The face is a web page (React + Vite + Three.js + custom GLSL). The brain is Claude Code, run headless as a library.

The only subscription you need is Claude Code. No API keys, no OpenAI account, no cloud bill — the brain runs on your existing Claude Code login, and the heavy work (the model itself) runs on Anthropic's servers, so even a low-end laptop only has to draw the interface. ElevenLabs is an optional add-on that gives JARVIS a much better voice and sharper hearing; without it he speaks and listens through the browser's own speech, and everything still works.

What this fork adds

• Bridge bound to loopback. The bridge used to listen on every network interface, and its Origin check is a browser convention, not authentication — any client on the same network could set the header and drive an agent that may have shell access. It now binds to 127.0.0.1 only (bridge/server.mjs).
• Windows support. The chrome_* extension bridge does not run on Windows, so on win32 the system prompt points JARVIS at the Chrome DevTools MCP server instead and teaches it the Git Bash quirks of Windows commands (taskkill //IM, winget). macOS keeps the original behaviour. The prompt also requires a spoken confirmation before shutdown, restart, sign-out or deleting files.
• "Jarvis, logout". A voice dismissal that acknowledges and returns to standby. The pattern must match the whole utterance, so "log out of my account" is still treated as a task rather than a dismissal (src/App.tsx).
• Unattended mode. ?auto=1 powers up on load, with no INITIALISE click or clap, so JARVIS can run as a background assistant.
• Windows launchers. Start JARVIS in a Chrome window parked off-screen (kept un-throttled so the wake word still works), bring it back into view, and stop exactly the processes it started. See Running on Windows.

Requirements

In one line: a Claude Code subscription, plus two free things every computer can have — Node.js and Chrome. That's the whole list.

• Claude Code, installed and logged in — this is the only account you need. Install it with the official method — npm install -g @anthropic-ai/claude-code, or the platform installer at https://docs.claude.com/en/docs/claude-code — then run claude once and complete login. The bridge reuses that login. No API key, and usage is billed to your existing Claude account.
• Node.js 20 or newer — free, one installer from https://nodejs.org. This is a Node web app, so it is the one unavoidable tool.
• Google Chrome or Microsoft Edge, in a real browser window — not an embedded preview pane. Preview panes (including the one inside editors and Claude Code) block microphone access, so the page loads and looks right but never hears you. JARVIS also needs WebGL, which these browsers provide.
• Optional: an ElevenLabs API key — a good add-on, not a requirement. It gives a better voice and sharper transcription; the free tier is plenty for a demo. Without it, everything runs on the browser's own speech.

Run npm run setup after cloning and it checks all of this for you, in plain language.

Quick start

First, install, then start it:

npm install
npm start          # runs the brain and the face together

Then open the URL it prints (http://localhost:5173) in Chrome, click INITIALISE, and say “Hey Jarvis”.

Prefer two terminals? Run them separately instead:

npm install

Terminal 1 — the brain:

npm run bridge

Terminal 2 — the face:

npm run dev

Then open the app in a real Chrome or Edge window:

open http://localhost:5173

Click INITIALISE, allow the microphone when asked, and say "Hey Jarvis".

It has to be a real browser window. Embedded preview panes block the microphone, so JARVIS will look perfectly alive and simply never respond.

Running on Windows

Double-click launchers live in the repo root:

Both start modes run with writes enabled — read Enabling actions first.

The hidden launcher uses a Chrome profile of its own (%LOCALAPPDATA%\JarvisApp) with the microphone pre-allowed for localhost:5173 and the camera denied, because nobody can answer a permission prompt in a window they cannot see.

For the ElevenLabs voice, copy jarvis-secrets.example.cmd to jarvis-secrets.cmd and paste your key in. That file is git-ignored.

For browser control, add the Chrome DevTools MCP server to your user-level Claude Code config, where the bridge looks for it:

claude mcp add --scope user chrome-devtools -- npx chrome-devtools-mcp@latest

How it works

JARVIS is two processes. The browser is the face and the voice; the bridge is the brain and the hands.

  ┌─ browser (the face) ───────────────┐        ┌─ bridge (the brain) ─────────────┐
  │  "Hey Jarvis" wake word            │        │  Node · bridge/server.mjs        │
  │  local VAD  →  speech to text      │   ws   │  Claude Agent SDK                │
  │  reactor UI (Three.js + GLSL)      │◄─────► │   = Claude Code, headless        │
  │  text to speech                    │  8787  │  spawns your MCP servers         │
  │  heads-up display                  │        │  permission gate (decideTool)    │
  └────────────────────────────────────┘        └──────────────────────────────────┘

Everything you see and hear happens in the browser. The bridge is a single Node process (bridge/server.mjs) that runs the Claude Agent SDK (@anthropic-ai/claude-agent-sdk) — this spawns the real claude CLI as a child process, so the brain literally is Claude Code, headless. They talk over a WebSocket (plus a few HTTP endpoints) on ws://localhost:8787.

Why a bridge at all? A browser tab cannot spawn the local stdio MCP servers — higgsfield, elevenlabs, android, playwright, exa, serper, and the rest. The bridge can. And because it is the Agent SDK, it authenticates off your existing Claude Code login: no API key, billed to that same Claude account.

The model. claude-opus-5 at effort medium by default. Override with the JARVIS_MODEL and JARVIS_EFFORT environment variables. On startup the bridge prints its choice, e.g. [jarvis] model claude-opus-5 · effort medium.

The voice pipeline

The loop is designed so that nothing silently dies and barge-in feels natural.

• Detection is local. An energy-based voice-activity detector (src/lib/vad.ts) decides when you are speaking. It is instant, cannot quietly fail, and is what makes barge-in work — speak while JARVIS is talking and he stops.
• Transcription has two tiers, chosen automatically at boot. The browser asks the bridge /health and picks the best available:
  • ElevenLabs key present → ElevenLabs Scribe, via the bridge /stt endpoint.
  • Nothing configured → the browser's own SpeechRecognition (Chrome/Edge), guarded by a heartbeat so it recovers when Chrome throttles it.
• Speaking uses the ElevenLabs voice when a key is present, and the browser's speechSynthesis otherwise. If a cloud call fails it falls back to the browser voice, and if the OS voice itself is broken it latches over to the cloud voice.

So it works with no keys and auto-upgrades when a key appears — there is no flag to set. Capability detection lives in src/lib/capabilities.ts, which probes the bridge's GET /health (returning { ok, tts, stt }, both tracking the ElevenLabs key) once at boot and picks the engines.

What JARVIS can do

Beyond answering, JARVIS reaches every MCP server in your Claude Code configuration, and can drive his own interface.

Your tools

Every server in your ~/.claude.json is handed to the SDK explicitly. Depending on what you have installed, that is roughly:

• Web & search — exa, serper, serpapi
• Images & video — higgsfield, openrouter-image, palmier-pro
• Voice — elevenlabs
• Your phone — android
• The browser — playwright

A few things you can say:

• "What's happening in AI this week?"
• "Generate an image of the Mark VII suit."
• "Take a screenshot of my phone."
• "Open my GitHub notifications."

Note on account connectors. Servers you added through your claude.ai account are not stored on disk, so the bridge cannot see them — it works from the servers in ~/.claude.json (about 14), not the claude.ai ones.

JARVIS controls the interface

He drives the UI through MCP tools the bridge exposes:

• ui_theme — accent, background, per-phase colours
• ui_reactor — colour, scale, intensity, spin, and style (ring | sphere | wire), visibility
• ui_orbit — put images in orbit around the reactor
• ui_chrome — show or hide rails, transcript, badges
• ui_effect — glitch | pulse | scan | shake | flash
• ui_screen — clear
• ui_reset — back to defaults

So "make it red, hide the systems list, put that render in orbit" is a spoken command.

The heads-up display

JARVIS authors panels with a display tool against a fixed .hud-* design system. The browser sanitises the markup (DOMPurify, a class allowlist and a strict CSP) before rendering. Rich media works — images, <video>, and YouTube/Vimeo embeds. Remote images and video are fetched server-side through the bridge (/img and /media, both SSRF-guarded), so hotlink-blocked news thumbnails still appear and the page never beacons your IP to a host the model chose.

Controls

The boot sequence

Power-up plays a four-beat Iron Man start-up (src/ui/Boot.tsx): an "INITIATING SYSTEM" status bar with a segmented progress bar and boot log; then concentric reticle rings resolving into "J.A.R.V.I.S"; then a suit schematic; then the triangular arc reactor lighting up — with a start-up sound under it (public/audio/boot-music.mp3).

Configuration

Everything is optional in bridge mode. Frontend settings live in .env.local (copy .env.example); bridge settings are environment variables.

Bridge

Frontend (.env.local)

Adding an ElevenLabs key

You do not have to touch a flag. Either:

• Set ELEVENLABS_API_KEY on the bridge before starting it, or
• Add the key to your elevenlabs MCP server's env in ~/.claude.json — the bridge reads it from there too.

Either way, /health starts reporting the capability, the browser picks it up on the next boot, and both the voice and transcription upgrade automatically.

Enabling actions

The tool gate starts read-only. Search, generation and lookups run freely; anything effectful — send, tap, delete, install, pay — is denied. Voice is a poor interface for a confirmation dialog, so the decision is made ahead of time in decideTool() in bridge/server.mjs, not at the moment of use. The bridge sets settingSources: [], which makes its own gate the only authority — filesystem settings and any global bypassPermissions cannot override it.

To allow effectful tools (phone, browser driving, sending), run the bridge this way instead:

npm run bridge:writes

Read decideTool() before you do. "Hey Jarvis, clean up my downloads folder" means something rather different with writes enabled.

Troubleshooting

I can't hear him, or he can't hear me. Press D for the diagnostics panel — it states plainly whether he is hearing you and whether he is producing sound. Press T for a one-line audio self-test.

No voice at all. You must be in Chrome or Edge, in a real browser window (not an embedded preview), and you must have allowed the microphone.

Bridge not reachable. Check that npm run bridge is still running in its terminal, and that nothing else is holding port 8787.

Security

All of this lives in bridge/server.mjs:

• The bridge listens on 127.0.0.1 only, so nothing else on the network can reach it. The origin check below is a browser convention, not authentication.
• The WebSocket accepts only local dev origins (add more with JARVIS_ALLOWED_ORIGINS).
• /file, /img and /media validate the scheme, confine to allowed roots, resolve the real path, and refuse private and loopback addresses (SSRF guard).
• The tool gate (decideTool) is default-deny for effectful MCP tools.
• A strict CSP in index.html; model-authored panel HTML is sanitised.

Credits & licence

MIT. The original project is by Aditya Dewaskar — adewaskar/jarvis. The additions in this fork are released under the same licence.

Music by Kevin MacLeod (incompetech.com), licensed under Creative Commons Attribution 4.0 — see public/audio/CREDITS.md.

The boot sound and any tracks in public/audio/ ship with the project for the demo. If you go on to monetise something built on this, clearing the rights to that audio is your responsibility.