# TIKOCI Open Source RouterOS Tool

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Anthony Shaffer](https://cerebralvalley.ai/u/Amm0)
- **GitHub:** https://github.com/tikoci/routeros-muse
- **Demo video:** https://youtu.be/dL-iZO-o6M8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4720

RouterOS-Muse puts an AI coding agent beside a real MikroTik network, with a workspace to learn it, tools to act on it, a disposable twin router to test on, and a browser UI where the owner watches, approves, and sees results.

The problem: coding agents can now read a codebase, use tools, and verify their own work. Networks get none of that. The facts that matter live on routers, and owners must re-explain their topology in every chat. A chat reply cannot log into a router, and nothing it learns lasts past the session.

The workspace is a folder that outlives every chat. workspace.toml says what the network is. An encrypted device.cdb (the WinBox-compatible registry) holds devices and credentials. devices.toml is a read-only mirror of the non-secret parts for plain file tools. router.toml records per-device intent. Observations, reports, operations, and history make up the rest. "Agents learn their job from a directory": the network template ships AGENTS.md, conventions, checks, and skills, so the agent arrives knowing how this network is run.

The agent acts through the TIKOCI RouterOS stack: Rosetta for documentation and command-tree research, Centrs for device access and validating commands before they run, QuickCHR for disposable CHR routers on QEMU, and routeros-skills for workflows. The rule: "Prove it on a twin. Apply it only with permission." Changes are tested on a CHR first, and Muse's approval modes gate anything that touches the real network. A CHR cannot prove radio or physical-port behavior, and the skills say so.

The UI is more than a chat box. Devices sit on the left, the conversation in the center, and renders and status on the right. A per-session MCP server gives the agent UI tools. ui_open puts a report, diagram, or agent-built HTML page in a tab. ui_select_device targets a router. ui_add_device opens the credentials dialog, so a router password is typed by the owner and never enters chat or argv.

It is real software, not a mockup. It is published on npm (bunx @tikoci/routeros-muse@next start). CI is green and runs lint, a typecheck, 461 tests, and a Playwright browser smoke. The repo has 328 commits in four days. Work is tracked in records/ (177 tasks, 55 decisions). Gaps found in the tools it consumes were fixed upstream, not forked.

Two builds show the range. enbeep was Muse's first assignment: a RouterOS :beep script that sends a router's MAC and IP as MFSK-16 tones, and a Web Audio decoder that hears them (tikoci.dev/enbeep). app-t3code, built from inside RouterOS-Muse, packages T3 Code as a RouterOS /app container for x86_64 and ARM64, tested on real CHR VMs.

Status: preview. Narrow-window layout and some UX edges are known issues.

## More from Meta Global AI Developer Hackathon

- [DB Creations LLC (Vibe Audit 2.0)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4717)
- [TheNocturnalCoder](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4718)
- [HomeWise AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4719)
- [DNAFix](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4721)
- [EHIBE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4722)
- [IA Para Todos](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4723)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4720. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt