TOML Agent Engine
An agent engine with no SDK, just TOML and webhooks I was looking for a way to add AI agents to multiple systems that I maintain. I had a few requirements: - I work in multiple programming languages, so it needs to work for all of them. - I can control how tools access my system. - I can override the agent loop to try different patterns. - I can iterate on the loop without handling durability, streaming, and the rest of the nitty gritty. - I can iterate locally before deploying. - I can self host. Nothing existed that worked across programming languages, so I built it. It's not an SDK. It's a server that handles durability, streaming, MCP connections, LLM calls and connecting to channels like Slack. You can override every single step of the agent loop by handling a webhook. But you don't have to override anything, you don't even need to handle the webhook. You can define the agent in a TOML file and chat with your agent via Slack in minutes. Point it at your backend to receive a webhook at every decision point. The engine sends the current step and a proposed next step. You can echo the proposal to do what the engine wants, or you can do something else and customize the agent loop. Check it out! https://github.com/substructureai/substructure https://substructure.ai/