Convex Agent Plugins
Convex publishes official plugins for the major coding agents. With the plugin installed, you can describe an app in one sentence and watch your agent scaffold it, running, in front of you, then keep shipping features while it reads your real deployment instead of guessing, catches its own mistakes as it works, and follows idiomatic Convex patterns.
This page explains what the plugins do, how to install them, how to get the most out of them, and how to send us feedback. For agent-specific setup, see the dedicated pages for Claude Code, Codex, and Cursor.
What the plugin does
Every Convex plugin bundles three kinds of help:
- Tools: a built-in Convex MCP server that lets the agent securely interact with your dev deployment. It can read your data, logs, and insights, and run functions, so the agent works from what your deployment actually contains instead of guessing.
- Hooks and monitors: background checks that run as the agent works. A pre-commit typecheck, an end-of-turn verify loop, and monitors that surface dev and production errors let the agent find and fix issues without you having to point them out.
- Skills and specialized agents: focused instructions that teach the agent how to do specific Convex tasks correctly, invoked automatically when relevant or on demand.
What you can do with it
A plain-English request goes a long way. The agent picks the right skill automatically, or you can invoke one explicitly:
| Ask for… | The skill | What it does |
|---|---|---|
| "build me a todo app" | quickstart | One sentence in, a running app out, scaffolded and built live in front of you. |
| "add sign-in" | auth | Full authentication wired end to end, passkeys by default, including the auth.config.ts that trips everyone up. |
| "add a chatbot" | agent | A real in-app AI agent backend: threads, message history, tool calls, and vector-search RAG. |
| "change my schema safely" | migrate | Change a live schema without breaking data: stage, backfill every existing row, verify, then tighten. |
| "add <a feature>" | add | Pulls in the right piece of the Convex component ecosystem: crons, file storage, and more. |
| "review my functions" | convex-reviewer | Audits your convex/ code for validator, auth, and performance issues and shows you the fixes. |
On top of the on-demand skills, an always-on Convex specialist quietly reviews
every function the agent writes under convex/, so idiomatic APIs, correct
validators, and indexed queries are the default, not something you have to ask
for. New skills land regularly.
Install
Pick your agent. Each install is one command (full walkthroughs on the per-agent pages linked above):
Claude Code/plugin install convex@claude-plugins-official
Codex (from Codex's built-in plugin directory)
codex plugin add convex@openai-curated
For the newest build straight from Convex, see the nightly marketplace install.
Cursor: install from the Cursor Marketplace, or manually:
cd ~/.cursor/plugins
git clone https://github.com/get-convex/convex-agent-plugins convex
# then restart Cursor
Using the plugins effectively
A few habits get you the best results:
- Describe the outcome, not the API. "Let users upload profile photos" routes to the right skill better than naming internal Convex primitives. The plugin knows the mapping.
- Let the checks work. The hooks catch typecheck and deploy issues on their own; you rarely need to paste errors back in.
- Ask for a review before you ship. "Audit my Convex functions" runs the security, auth, and performance pass over your backend.
- Trust the deployment tools. For "what's in my database?" or "why is this slow?", the agent can read your dev deployment directly through the MCP server instead of guessing.
Giving feedback
We're constantly working on improving the plugins with rigorous evals and real-world reports. You can help:
- Report a bug or request a skill by opening an issue on the plugin repo for your agent: Claude Code · Codex · Cursor.
- Improve the AI rules: if the agent writes non-idiomatic Convex, contribute a case to the convex-evals repo so we can measure and fix it.
- Ask the community in the Convex Discord.