routeros-muse

A RouterOS network workspace plus web UI that puts an agent beside your network. You keep your network's docs, devices, and history in one workspace folder; routeros-muse checks the machine, grounds the agent on your real routers, and gives you a browser UI to chat with it, watch it work, and view what it renders.

Built on the tikoci trilogy: quickchr (disposable CHR twins), centrs (device access plus credentials), rosetta (RouterOS knowledge). Started as a hackathon demo; now growing into a real owner tool.

Status: preview. Expect rough edges; nothing here is finished.

Prereqs: Bun (per Bun's own install instructions), plus the Muse agent from https://dev.meta.ai (muse login once to sign in; routeros-muse code checks and offers the installer). Then:

bunx @tikoci/routeros-muse start   # web UI on 127.0.0.1 by default; opens your browser

From a checkout, bun install once, then bun run start does the same. Creating a workspace asks for a CDB password — the device registry (device.cdb beside the workspace) is created encrypted, and the browser lands on the lock screen where you retype it. On a terminal start walks you through template + password. For scripted creation, pipe the password to routeros-muse init <dir> --password-stdin, then run routeros-muse start <dir>. The web picker's Create button asks in the form. --open-cdb creates without encryption (dev escape, never the default). make gate runs the full check (lint + tests + browser smoke).

Other commands: init scaffolds a workspace where you are, code checks the machine then drops you in a muse shell in the workspace, and doctor runs the same checks print-only.

A workspace is one folder that holds everything about a network: workspace.toml (what this network is), device.cdb (its devices: the centrs registry, mirrored read-only into devices.toml so plain text tools see hosts and ports), router.toml (per-device intent), operations/ and inbox/ (pending work and handoffs), plus observations, reports, and history. The agent reads the workspace; the workspace records what the agent did.

templates/network/ is the full starter template: conventions, skills, checks, and layout. New workspaces land in ~/Networks, one predictable path on every OS, and serving a workspace registers it for the no-workspace picker. The web UI's whole-app lock is your CDB password (see docs/design/t39-auth-design.md).

package.json pins @tikoci/quickchr, @tikoci/centrs, and @tikoci/rosetta, all three on the next tag during preview (this checkout's lockfile resolves centrs to 0.1.9; Status shows the installed versions). The bootstrap and the server front the package's own node_modules/.bin on PATH for every spawn — server subprocesses and agent children alike — so all sides resolve identical builds:

• quickchr — twin capability probe (quickchr doctor) and disposable CHR routers for grounding claims, configs, and scripts.
• centrs — LAN discovery (MNDP), device reads and live stats, and the CDB credential store (device.cdb beside the workspace).
• rosetta — RouterOS device database and knowledge lookups for the agent, resolved from the same bundle.

routeros-muse [command] [dir] [options]   # bare = start
routeros-muse start [dir] [--port N] [--host H] [--no-open] [--yolo]
routeros-muse init [dir] [--template minimal|network] [--password-stdin] [--open-cdb]
routeros-muse code [dir] [muse args...] [--yes] [--router H[:port]] [--print-only]
routeros-muse doctor [dir]

New workspaces are created encrypted: init and start option 1 prompt for the template (--template skips) and a CDB password (hidden, typed twice) on a TTY, or take it from --password-stdin off a TTY. --open-cdb is the explicit escape to an open registry.

start (alias: serve; bare means start): web UI over the workspace — devices left, chat center, render/status right — opened in your browser; --no-open prints the URL instead. Defaults: host 127.0.0.1 (--host overrides the bind; only do that on a trusted network), ephemeral port, provider meta. When the directory (default: here) is not a workspace, a terminal offers init-here / registered-workspace pick / browser picker; off a terminal the default (here) serves with no workspace attached, while a named directory that is not a workspace stops with the init line. --yolo starts the agent session at its most permissive approval mode (Allow all). The agent worker spawns trusted by default, so the workspace's AGENTS.md and .agents/skills/ load (Settings shows the state). GUI sessions also load the packaged RouterOS skills by default, including without a workspace or with workspace trust off. In Settings, Automatically load RouterOS skills changes this for every GUI session on the machine. The change applies to the next new or resumed chat; it never interrupts a running turn. Independently installed or workspace skills can still load. The skill version is pinned with the package: startup works from local files, without contacting GitHub. --tui aliases code. Developer flags: --no-workspace, --provider NAME, --allow-open-cdb (an open CDB implies no auth), --no-trust-workspace (spawn the agent untrusted: the workspace's AGENTS.md and skills don't load).

Settings → Disable sandbox enables direct networking while retaining the selected approval mode, including Ask first. It removes filesystem and network isolation for web UI agents on this machine; approved commands can access paths outside the workspace and any reachable address. Host allow rules do not restrict those connections. Default: off. After changing it, finish the current turn, stop routeros-muse and start it again with the same workspace. Settings shows when a restart is pending. A page reload or New Session does not apply it; reopen the previous chat from Past chats.

init: scaffolds a workspace in the directory (default: here) from minimal (runnable minimum) or network (full starter), never overwriting, with an encrypted device.cdb, and registers it. init then exits, printing the start command; start opens the browser on the lock screen, where you retype the password you set. The web picker's Create button scaffolds into ~/Networks instead, then navigates there itself.

code: checks prereqs, the trilogy, and the agent; opens and validates the workspace; reconciles [pins]; probes twins and LAN into a capability mode (full / twin-only / real-box-only / design-only); lists handoffs; then execs muse in the workspace with the package tool bundle on PATH. --yes is the one-yes for echoed user-level installs; --print-only prints the launch line instead of running it; every other arg passes through to muse verbatim and in order (-c, --yolo, resume flags), with -- forcing the rest through. The workspace directory, if you name it, goes first, before any muse arg. Nothing installs without asking.

doctor: the same checks print-only. Never installs, never execs muse. Exit 1 when a required item is missing.

Fleet reads live in centrs (devices list, retrieve), not here; agents use centrs/quickchr/rosetta directly.

• SECURITY.md — security facts and known gaps.
• docs/internals.md — how the pieces fit (bootstrap steps, server routes, render services, unlock, registry, smoke).
• docs/design/ — design notes carried from the staging archive.
• docs/concepts/ — UI concept mockups (design references, not product).
• records/ — the build tracker (needs/tasks/decisions/concepts).
• templates/network/ — the starter workspace template.
• tools/ — skill gate, UI smoke, router capture plus checker.