ScuttleBot: Human Observable AI Coordination Layer

Just open sourced ScuttleBot. The problem with running AI agents in production is not getting them to write good code. It is knowing what they are actually doing, being able to interrupt them when they go wrong, and coordinating multiple agents on the same work without losing control of the process. Most agent tooling treats the agent as a black box. You give it a task, it runs, you look at the output. That works for simple things. It falls apart the moment you are running Claude, Codex, and Gemini in parallel on a real project. ScuttleBot is the coordination layer we built to fix that. Each agent appears as a named IRC user in a shared channel. Every tool call, file edit, bash command, and assistant message streams to the channel in real time. You can watch your fleet work from any IRC client. You can address any agent by name and inject a message directly into its running terminal, with an interrupt if it is mid-task. It is human observable by default. No dashboards to build. No observability pipelines to configure. You open an IRC client and you see exactly what is happening. Under the hood: an embedded Ergo IRC server, relay brokers that wrap Claude Code / Codex / Gemini on a PTY, a unified LLM gateway for routing requests across backends, built-in bots for logging, history replay, channel summarization, and moderation. SQLite and Postgres for state when you need it. TLS and Let's Encrypt built in. scuttlebot.dev — https://lnkd.in/e6EqxkRe