Convex Sub-Agent for Claude Code
Hey all,
I made a basic config for a sub-agent for Claude Code leveraging the existing official rules provided by Convex. Hope this can help someone else get off to a good start using the new agents feature! Open to PRs if anyone sees ways to improve this!
https://github.com/mwarger/convex-agent-config
Post here: https://x.com/mwarger/status/1948780703589544247
GitHub
GitHub - mwarger/convex-agent-config: Convex backend development ag...
Convex backend development agent configuration. Contribute to mwarger/convex-agent-config development by creating an account on GitHub.
Mat Warger (@mwarger)
I used the anthropic rules supplied by @convex_dev to make a Claude Code sub-agent config. I used this for a couple things and found it to nice encapsulate the rules. This seems so far to work better than having the rules in context or mentioning them explicitly. Link below:
X
8 Replies
Great ideer! 🙏🏽👏🏽
Hey @mwarger , how has your experience been with claude code and convex?
Literally did a search to see if anyone had done this haha. Have you considered including some stuff from convex-helpers as well such as relationship helpers? @mwarger
Those are great ideas - I recently updated the repo to split the agents out. I'm still playing with what might work best.
I have had good experiences with it so far. I have had a pretty rough experience with testing.
Like any context engineering, it's best to have good patterns and examples you can use to show it how to work effectively.
Give it links to the documentation and ask it to generate the examples based on it. I was planning to make one big agent, but I'll try your setup.
Convex components are another use case too
I agree. Part of what I learned after looking further into the agents documentation was that the way you structure the prompt for the agents matters. It seemed easier to do this when they are more focused. They can pass smaller tasks back to the primary agent once they're split up. As one trade-off, they need to be more specific so Claude can effectively choose which agent to use. But then again, the new ability to mention agents explicitly is a big help in that regard. You can then effectively have pre-made prompts that you can hand work too and they do specific things. This level of orchestration could become pretty powerful. It would be neat to make this into a sort of state-machine style workflow.
Yeah this is very exciting, can’t wait to try it later this week. Definitely +1 on adding the convex-helpers, relationship helpers, and components… when OpenAI launched CustomGPTs I saw a few people create them for convex, but never got around to using them, hoping this becomes a better answer…
This is amazing, it opens up even more streamlined way to use MCP Agents not only with Claude Desktop but you can do the same with Cursor, Windsurf and VSCode to name a few more...