mona-agent — your computer, with a brain in the cloud

The open-source AI agent for your computer.
Chat from anywhere. It runs commands, manages files and automates tasks —
private by design, with your AI keys locked in an encrypted cloud vault.

Contents

• Private & secure by design
• What it can do
• Built for
• Quickstart — 60 seconds
• How mona-agent compares
• Documentation
• Security & compliance

---

The story

Most AI assistants live in a chat window. They can write you a script — but they can't run it. They can explain a crash — but they can't look at your logs. They know everything about your code and nothing about your machine.

mona-agent is the other half. It's a small, fast daemon that lives on your computer — macOS, Linux, WSL2 — and connects it to a brain in the cloud. You talk to it from your dashboard or your phone; it reasons about the task, then actually does it: checks disk space, restarts a service, finds a file, opens an app, runs a cleanup. And it streams every step back so you can watch it think and act in real time.

It's the difference between an advisor and an employee. One tells you what to do. The other just gets it done.

  Your Device                           mona.expert Cloud
  +-------------------+                +----------------------------+
  |   mona-agent      |                |  Dashboard & chat          |
  |                   |--- HTTPS ----+ |  AI engine (the brain)     |
  |  * executes tools |<-- commands --+ |  Agent orchestration       |
  |  * streams metrics|               |  Key vault (AES-256)        |
  |  * terminal UI    |-- results --->+ |  Live device monitoring    |
  +-------------------+                +----------------------------+

Private & secure by design

Your machine is a trust boundary, not a sandbox afterthought.

• Zero AI keys on your device. Provider keys (OpenAI, Anthropic, …) live only in the cloud, encrypted AES-256-GCM. The device holds one revocable token — nothing else worth stealing.
• Egress-only networking. The daemon makes outbound HTTPS connections and listens on localhost only. No open ports, no public exposure — it works behind NAT and firewalls.
• Guarded execution. Commands run through an allowlist; dangerous patterns (rm -rf /, fork bombs, silent sudo) are blocked before execution. File access is confined to safe paths with symlink-escape rejection.
• Complete audit trail. Every action — reasoning → tool call → result → answer → verification — is traced with tokens, cost and latency, and exportable for audits.
• Documented, not hand-waved. A real threat model (STRIDE), data-flow & minimization, security self-audit guide and EU CRA readiness. No fake certifications — "trust me" isn't a security model.

See SECURITY.md for the full policy and responsible disclosure.

What it can do

• Operate your computer from anywhere — ask from your phone: "how's the disk looking?", "restart nginx", "open Safari"
• Think act observe deliver — a real agentic loop: the cloud brain plans, the device executes local tools, results flow back, and the answer lands in your chat — up to 8 tool steps per task
• Fail is never allowed — transient errors retry automatically, failed commands trigger self-diagnosis and a smarter second attempt, and every task ends with an answer
• Zero secrets on the device — no AI provider keys are stored locally; they live in the encrypted cloud vault. One key. One brain.
• Live device monitoring — CPU, memory, disk, load and uptime streamed to your dashboard every 10 seconds, with history sparklines
• Terminal-native — a fast TUI with live logs, or a fully headless daemon; one command installs it, one dependency powers it
• Free and open source — MIT licensed, forever

Built for

• Developers — delegate test runs, log checks and service restarts to a device you can chat with from your phone.
• Homelab & server ops — manage headless boxes, Raspberry Pis and NAS without SSH; the device polls the cloud, so no open ports are needed.
• Personal automation — open apps, check disk, control media, run scripts from any browser.
• Kiosks & presentation machines — restart demo windows and long-running GUI programs on demand.
• Compliance-heavy teams — every action traced and exportable; EU CRA / NIS2 / AI Act / GDPR documentation included.

More scenarios: docs/USE-CASES.md.

Quickstart — 60 seconds

# 1. Install (needs Node.js 20+)
curl -fsSL https://agent.mona.expert/install.sh | bash

# 2. Log in with your mona.expert key
mona-agent login

# 3. Start the terminal dashboard — or run headless
mona-agent gui
mona-agent start

Then open https://agent.mona.expert/dashboard — your device appears with live stats, and you can chat with it from the browser.

Usage

mona-agent login            # save your mona.expert key
mona-agent connect          # test the connection end to end
mona-agent gui              # terminal dashboard with live log
mona-agent chat "free up disk space"     # one-shot conversation
mona-agent exec "uptime && df -h"        # run a guarded command
mona-agent start            # headless daemon with auto-reconnect

In the dashboard chat, just talk normally:

"Run df -h and tell me my disk usage." · "What's the load average?" · "Open Spotify and play some focus music."

The brain picks the tool, the device runs it, you get the result.

Built-in tools

Tool	What the agent can do with it
sysinfo	CPU, memory, disk, load, uptime, host and platform details
shell	Run commands — allowlisted by default, full shell opt-in
files	List, read, write, move, delete — confined to safe paths
net	HTTP(S) requests, DNS, connectivity checks

Full reference: docs/TOOLS.md

How mona-agent compares

	Typical AI chat	mona-agent
Can it see your machine?		live metrics & files
Can it execute?		sandboxed shell & tools
Does it retry when things fail?		auto-debug + retry loop
Where do your API keys live?	on your disk	encrypted cloud vault
Is every action audited?		full trace, exportable
Install	app + account	one command, one dependency

FAQ

Is mona-agent free? — Yes. The client is MIT licensed and free forever; the cloud has a free tier at agent.mona.expert.

What data leaves my device? — Only what the task requires: system metrics, and the results of commands you asked the agent to run. Nothing is sent to third parties. See docs/GDPR.md.

Can it damage my machine? — Commands are allowlisted by default, the file tool is confined to safe paths, and every action is recorded in the audit log. Full shell execution is an explicit opt-in.

Does it run on servers? — Any Node.js 20+ machine: headless Linux boxes, Raspberry Pis, home servers. See docs/GETTING-STARTED.md.

Documentation

Page	Contents
docs/GETTING-STARTED.md	Install, login, first steps, troubleshooting
docs/ARCHITECTURE.md	Daemon internals, agentic loop, metrics pipeline
docs/TOOLS.md	Tool-by-tool reference with examples
docs/EXAMPLES.md	Automation recipes — cron, watchdogs, backups
docs/INTEGRATIONS.md	Cloud API, scheduling, boot persistence
docs/FAQ.md	Frequently asked questions
docs/CHANGELOG.md	Release history
examples/	launchd & systemd units, health check script

Compliance & Trust

• EU Cyber Resilience Act — SBOM, vulnerability handling, secure by design, coordinated disclosure
• NIS2 — TOMs, logging, incident support
• EU AI Act — limited-risk classification, transparency obligations implemented and documented
• GDPR — data minimisation, Art. 30 record, DPA-ready
• SBOM — CycloneDX 1.5, one runtime dependency
• Security policy — threat model, disclosure process, supported versions

Community

Star the repo if mona-agent is useful to you — it genuinely helps others discover the project.

• GitHub Discussions — ideas, questions, show & tell
• Issues — bugs and feature requests
• Hugging Face — mirror of this repo
• agent.mona.expert — the cloud dashboard
• SECURITY.md — responsible disclosure

Development

git clone git@github.com:MONAEXPERT/agent.git
cd agent
npm install
npm test            # 26 tests, 9 suites
npm run gui         # run the dev build of the TUI

Requires Node.js ≥ 20. Plain modern JavaScript (ESM), no build step, one runtime dependency — deliberately small.

License

MIT — see LICENSE. Free forever.

---

mona.expert — one key, one brain, any device.

Compliance & standards

Enterprise-ready documentation: EU CRA readiness, ISO/IEC 27001 mapping, IEC 62443 alignment, threat model, data flow & minimization, deployment guide, enterprise FAQ and SBOM. See SECURITY.md for vulnerability reporting.