Hey all,
I wanted to share a project I’ve been building called Odin:
Website: https://www.odin-bot.net/
Github: https://github.com/Calmingstorm/Odin
Odin is a self-hosted Discord-native AI execution agent. Instead of just chatting, it can actually use tools: run commands on configured hosts, work with Git/Docker/Kubernetes/Terraform, inspect logs, read files, browse websites, take screenshots, schedule checks, generate reports, and validate services after changes.
It can also handle larger development workflows end-to-end: scaffold or modify a project, write code, run tests, compile/build it, package artifacts, and post generated files or binaries back into Discord. It’s not limited to infrastructure checks — it can do real coding work too, with the same execute/verify/report loop.
Example: I’ve used Odin for full Minecraft mod development, end to end. It created and iterated on a NeoForge mod, built the working `.jar`, launched a local Minecraft client/server test environment, and validated gameplay behavior including block interaction and the mod GUI. It also produced project documentation and pushed the source to my GitHub repository.
That all came from high-level prompts in Discord, not me manually babysitting every file and command. Screenshots of progress from in-game were posted to Discord as he progressed through development.
I built this because I was tired of essentially begging OpenClaw to actually proceed, and do the things it talked about being able to do, but needing constant pushing only to deliver incomplete projects.
Odin also has a set of guardrails designed to make it behave less like a slippery chatbot and more like an execution agent. Each Discord request is separated from prior history with a current-request boundary, so old context does not silently become a new task queue. Its response guards check for fabricated action claims — if Odin says it checked logs, ran a command, verified something, created a file, uploaded an artifact, or inspected a system without actually using a tool, the response is rejected and retried with a tool call. It also catches promise-without-action phrasing, permission-seeking hedges, shell commands pasted instead of executed, false “tool unavailable” claims, and premature failure after only one broken attempt. After tools run, a completion classifier can decide the task is still incomplete and force up to three continuation passes. The goal is simple: Odin should execute, verify, and report real results instead of asking for unnecessary permission, stopping halfway, or inventing receipts like every other agent I've worked with.
It's free, it's opensource - I don't care if you do / don't use it, just wanting to share because it's been helpful to me, and was built out of frustration with existing solutions. It's probably not perfect, but a lot of time was spent on it and I think it's in a solid state. Give it a try, or don't - up to you, if you need help reach out, open an issue - I try to keep the website updated for install instructions / docs. .deb releases for easy installer, built for Linux.
My Cool Project Post #8000