OpenClaw
A self-hosted AI gateway. One Node.js service owns the model connection and serves the same assistant to Discord and Telegram.
The problem
An assistant spread across a browser tab, a phone app and a chat window is really three assistants. I wanted one service that owns the model connection and the configuration, runs on hardware I control, and answers wherever I already am.
What I built
- The gateway. A Node.js service that holds the provider connection and reads one JSON config file.
- Jarvis on Discord. The main interface, because Discord is where I already spend my time.
- A Telegram adapter. Added second, to prove a new surface doesn't touch the core.
- Readable failures. When something breaks, the log says what and where.
How it works
Fast, but rate limits made it unreliable for something always on.
Hit the same wall from a different direction.
Nemotron through NVIDIA's hosted API. Stable enough to forget about.
Pay-per-use access to Claude, so cost follows actual use.
Two providers failing the same way told me reliability was the real design constraint, not model quality. The provider sits behind a boundary, so every switch was a config change and neither bot needed rewriting.
Next up: separate Dev and Biz agents as their own Discord bots, GitHub integration so the assistant can act on repositories, and finishing the OpenRouter cost comparison.
What I learned
- Rate limits and provider reliability shaped the architecture more than model choice did.
- Draw the boundary before the first migration and the second one is cheap.
- Self-hosting trades convenience for control. Make that trade on purpose.
- Tool use, streaming and prompt structure came straight from the Anthropic Claude API course.