MOSS — Managed Operations & Systems Service

Your AI IT department, running on your own hardware.

MOSS looks after your home or small-office network with a team of AI agents. You hire them (a Network Admin, a Systems Admin, a Security Admin) and they get to work: finding every device on your network, keeping an eye on the services you care about, investigating when something breaks, and suggesting fixes.

Nothing changes on your network without your say-so. Agents work inside the budgets, permissions and network boundaries you set, any change they want to make waits for your approval, and every action is recorded.

Early release (v0.5). MOSS works end to end, but it's young software: expect rough edges and changes between releases. Only point it at networks you own or are allowed to manage.

What MOSS does

• Knows what's on your network. Agents scan the networks you allow and keep an inventory of every device: what it is, its address, who made it. Lock a device and agents can't change its inventory entry.
• Watches your services. Ping, web, port, certificate and DNS checks, or alerts from tools you already run (Uptime Kuma, Beszel, Prometheus Alertmanager, or any script). Build dashboards, and put one on a TV.
• Works the problem. When something goes down, MOSS opens an incident and an agent starts investigating straight away. When it comes back, the incident is updated.
• Fixes things, with your approval. An agent that wants to change something (restart a service, reboot a server, block a device) writes a change request: the exact steps, and how to undo them. It waits for you to approve, from MOSS or from your phone.
• Agents you can talk to. Chat with any agent, give it a task, or ask Moss, the agent who knows MOSS inside out. Each agent has its own skills, schedule and spending limit, and the Basement shows who's working and who's on a break.
• Use any AI model. Anthropic, OpenAI, OpenRouter, any OpenAI-compatible service, a Claude Pro or Max subscription, or a local model with Ollama so nothing leaves your network.
• One switch stops everything. The kill switch pauses every agent immediately.

Integrations

MOSS works with what you already run: Home Assistant, Proxmox, Docker, Pi-hole and AdGuard Home, OPNsense and pfSense, your UPS, Tailscale and Cloudflare, with notifications and approvals in Telegram or Discord. Each one adds health checks, fills in your inventory, and gives agents safe, approved ways to help.

For businesses, MOSS adds single sign-on with Microsoft or Google, Microsoft 365 and Google Workspace, Slack, PagerDuty, FortiGate, Cisco Meraki and AWS (these need a business licence).

See every integration and what it does →

How it keeps you safe

MOSS assumes the AI can make mistakes, and that your network can contain hostile data (a device named "ignore previous instructions" is a real attack). Safety never depends on the AI behaving.

• No shell. Agents can only use specific, checked tools (scan this network, check this port). None of them runs arbitrary commands.
• A gate checks every action: the kill switch, the agent's budget and permissions, and whether the network is one you've allowed. Networks you mark off-limits are always refused.
• Changes need approval. Anything that changes a system must match a change request a person approved, step for step.
• Secrets never reach the AI. Passwords and tokens are encrypted, and agents only ever see a name like secret:router-ssh. Each secret works only on the devices and tools you allow.
• Untrusted data stays data. Device names, scan results and webhook messages are passed to agents as information, never as instructions.
• Everything is recorded in an audit log you can verify hasn't been altered.

Want more? The hardening guide explains the optional settings: keeping agents and the toolbox off the internet, device key pinning, two-factor approvals, stricter chat approvals, agent wiki drafts, backups that never leave the host, and more, with what each one protects against and what it costs. Using MOSS at work? Read its first section, At work: start read-only.

Quick start

You need a computer that stays on (Linux is recommended), Docker with the Compose plugin, and openssl. MOSS is tested on x86-64; ARM boards such as the Raspberry Pi 5 are planned but not yet tested. On Windows, run these commands in WSL or Git Bash, not PowerShell.

git clone https://github.com/Dan5672/MOSS.git
cd MOSS
git checkout "$(git tag --list 'v*' --sort=-v:refname | head -n 1)"   # the newest release
sh deploy/install.sh   # creates the settings and keys, then builds and starts MOSS

The first start builds MOSS on your computer, which takes a while: 10 to 20 minutes on a typical mini PC. When it's ready, install.sh prints a link like https://192.168.1.20/setup?code=…: open it and create your account. The code in the link means nobody else on your network can claim MOSS before you do. If the link says localhost, MOSS couldn't find your computer's network address: open the link on that computer, and add the address in Settings → HTTPS. Then:

1. Networks: add your home network (for example 192.168.1.0/24) and mark it allowed.
2. Agents → Models: add an AI provider (an API key, or a local Ollama address).
3. Agents: hire a Network Admin and let it discover your network.
4. Monitoring: add checks for the services you care about, and pick an agent to respond.

A checklist on the dashboard walks you through it. Back up deploy/secrets/master.key: without it, stored passwords can't be recovered, even from a backup.

New to MOSS? The Getting started guide goes step by step, and covers what to think about before giving an AI agent access to your devices.

Using MOSS from your phone and other devices

It works straight away: any device on your home network can open the address install.sh printed. There's no reverse proxy to set up and no ports to open on your router.

• The browser warning, once per device. MOSS makes its own certificate, so nothing is bought or exposed to the internet. Settings → HTTPS shows how to trust it on Windows, macOS, iPhone and Android.
• A different name or address? Change it in Settings → HTTPS, for example after your computer gets a new IP address.
• From outside your home? Use Tailscale. Don't open a port to MOSS.

More you can do

• Your own tools. Describe an HTTP API in a few lines of YAML and agents can use it, with the same checks as built-in tools. See Agents → Custom tools for an example.
• Run agents on a Claude subscription instead of an API key: add a provider of type Claude subscription in Agents → Models with a token from claude setup-token. MOSS strips out Claude Code's own tools, so agents only ever use MOSS's checked ones. Runs count against your plan's usage limits; using a personal plan this way is your decision under Anthropic's terms.
• Two-factor sign-in, a password policy, and roles for everyone you add.

Backups and upgrades

Everything lives in one database volume plus deploy/secrets/ and deploy/.env. You can also make backups and check for updates from Settings in MOSS.

sh deploy/backup.sh                          # back up the database, secrets and settings
sh deploy/upgrade.sh                         # back up, move to the newest release, and check it came back
sh deploy/upgrade.sh --rollback              # go back to the version before the last upgrade
sh deploy/restore.sh <backup>                # restore a backup (also how you move to a new machine)

The new version is built while the old one keeps running, so a failed upgrade changes nothing. Backups include the master key: copy them somewhere safe, off the machine.

Roadmap

• Next: passive discovery, a network map (LLDP and topology), email notifications, and full LAN discovery from the host.
• Later: a public integration SDK and MCP tool packs.

For developers

MOSS is a TypeScript monorepo (pnpm + Turborepo).

web       Next.js UI
worker    agent runtime, schedules, events, monitor scheduling
gate      the policy gate and secrets broker: the only service that uses the master key
toolbox   typed network tools (scans, probes, SNMP, SSH checks, device APIs), no shell
bridge    Telegram, Discord and Slack
postgres  PostgreSQL 17 + pgvector, and the job queue
backup    MOSS backups (reads the master key to archive it; hands archives out only encrypted)

apps/ holds the services, packages/ the shared code (database, domain logic, policy, tools, models, agent runtime) and library/ the agent templates, skills and Moss's own docs, as YAML and Markdown.

corepack enable
pnpm install && pnpm -r run build
# Tests need Postgres on localhost:5432: add docker-compose.dev.yml to MOSS_COMPOSE_FILES in deploy/.env
MOSS_TEST_DATABASE_URL=postgres://moss:<password>@localhost:5432/moss_test pnpm -r run test
cd apps/web && MOSS_TEST_DATABASE_URL=... pnpm exec playwright test   # browser tests

deploy/lab/compose.lab.yml adds a safe test network (an allowed subnet and an off-limits one, with target machines) and a scripted model, so you can try everything without an API key.

Contributing

Issues and pull requests are welcome. For anything bigger than a small fix, please open an issue first. Contributions need a Contributor License Agreement, because MOSS is offered under both an open-source and a commercial license. Please report security problems privately, not in a public issue.

When you change what people see or what MOSS does, update library/docs (Moss answers from it), run corepack pnpm gen:reference, and add a line to library/docs/changelog.md.

License

MOSS is open core under the GNU AGPL v3.0. Business integrations need a business licence, and a commercial license is available.