What is routerclaw?

RouterClaw is an autonomous AI agent written in Go, built to run on old OpenWrt routers with as little as 32MB of RAM and 4MB of flash storage. The target hardware is something like a TP-Link WR740N, a consumer router from around 2010. The idea is to turn a cheap, obsolete piece of networking equipment into an always-on smart home controller that sits on your network and responds to commands. You talk to RouterClaw through a Telegram bot. It can read and send Gmail, check Google Calendar, create and delete calendar events, scrape websites, submit web forms, and run web searches. It can also execute shell commands directly on the router itself, which lets it automate network tasks, monitor connected devices, and write its own scripts. A Wake-on-LAN feature lets it power on computers on your local network. It keeps conversational memory across interactions, and it can break complex requests into sub-tasks, chaining multiple tools together to reach a goal without human intervention. The technical setup is demanding. Because the router has only 4MB of storage, the standard OpenWrt firmware is too bloated to leave room for the agent. You have to build a stripped-down OpenWrt image that excludes the web interface and IPv6 support, then flash it by soldering serial pins to the router's motherboard and interrupting the boot process over a UART connection. The Go binary is cross-compiled for the MIPS architecture with software floating-point emulation, then copied into the router's RAM-disk filesystem at /tmp. A Windows batch script is included to automate recompilation and redeployment after power loss. The README is frank about the risks. RouterClaw can run arbitrary shell commands on its host, and while protections exist to stop it from accidentally rebooting the router, the author warns against using it on production networks. The project is experimental, and the README notes that new updates are coming.