Agentjarvis — Development skill for Claude Code

Agent Jarvis

Control **your own** Claude Code (or Codex) from your phone: chat with it, and approve or deny what it wants to do, from anywhere.

• A small host program, jarvis, runs on your computer (macOS or Linux) next to your agent.
• Your phone pairs with it once by scanning a QR code and typing a 6-digit code plus a passphrase only you know.
• Phone and computer talk end-to-end encrypted (Noise protocol) through a blind relay that only forwards ciphertext.
• When the agent asks for permission, your phone shows a card; every decision is signed by the phone and logged on your computer. No answer in 120 seconds = denied.
• The agent runs with your own login, settings and permission rules; on Linux it runs fenced (bubblewrap) so it cannot see or change Jarvis's keys or the list of paired phones.

**Zero access is the product rule:** our servers never hold a key that can read your messages, your agent's replies or your credentials. What we can and cannot see — and what is not done yet — is listed honestly, item by item, at .

Install

Tell your AI agent: **"read https://alpha.agentjarvis.net/install.md and do it"** — it installs, sets up and stops for you at every step only a human can do. [`install.md`](install.md) in this repository is the same file, byte for byte; the site publishes its SHA-256 at .

The host CLI on its own:

uv tool install "git+https://github.com/preangelleo/agentjarvis#subdirectory=host"
jarvis --version

What is here

Path	What
protocol/	wire spec (PROTOCOL.md), the shared JavaScript Noise implementation, official test vectors
host/	the host CLI jarvis (Python 3.13): pairing, relay connection, agent bridge, approvals, fence
web/	the phone web client (static, no dependencies) and its Cloudflare Worker
install.md	the instructions an AI agent follows to install Agent Jarvis

T