Here’s a little something I’ve been working on:
Hephaestus - an (opinionated) MCP server for AI agents working with Anvil apps:
Also mirrored at github but radicle is the canonical version and that’s where issues and patches live.
10 Likes
This is AMAZING!!
Have not tested yet, but will do soon and this is SO NEEDED thank you so much <3
2 Likes
Hey @owen.campbell I saw that this was moved to read-only - are you still investing in this project? It’s really cool.
I’m not maintaining it any longer. I’ve moved away from MCP as my preferred mechanism.
Interesting - may I ask what you’ve moved towards?
A couple of things caused us to move on…
-
context overhead - with an MCP server, every tool definition gets injected into every call whether it’s relevant at that point or not. For something like Hephaestus with a big app behind it, that’s a lot of tokens spent on context the agent often doesn’t need.
-
Around the same time, we started thinking about what would actually make our agents more effective generally and realised the answer was institutional knowledge. What if an agent could draw on everything we’ve learned over time, and add to that knowledge as it worked?
So we built a knowledge base that captures patterns, pitfalls, architectural decisions, that sort of thing and the agents search it as needed. We’ve built custom agents and a CLI around it too. Planning, execution, knowledge capture all have their own workflows. It’s taken a fair bit of work to get right, but it means the agents benefit from hard-won learning straight away. e.g. If we worked through a tricky Tabulator issue six months ago, that learning is available to any agent that needs it today.
It’s not something we’re planning to open-source but happy to have a chat if you’re exploring anything along those lines.
1 Like
Yeah, it’s very similar!