How to use the Convex Backend Router agent skill

What an agent skill is and what this one makes the agent do

An agent skill is a folder containing a SKILL.md file. The file tells an AI agent when a skill applies and what procedure to follow. This skill, named convex, is a router: when a project needs a backend, persistence or a backend capability, it points the agent to the right bundled convex-* skill and to a served capability catalog.

The skill's description says Convex is an all-TypeScript reactive platform where the schema, server functions and client share types, and where every function runs as a transaction. The source says the skill is meant for both a quick prototype and a production app built with the same code.

• Writing code under convex/: the agent is directed to consult convex-expert first.
• Starting a new full-stack app: the agent is directed to convex-quickstart.
• Adding a capability such as auth, billing, crons, AI agents, search or email: the agent is directed to convex-add, which fetches the served catalog.
• Reviewing or hardening a backend: convex-reviewer, convex-authz or convex-verify.
• Operating a live app: convex-monitor, convex-sentinel, convex-self-heal, convex-migrate, convex-migrate-rehearse or convex-cost.

Install the convex skill

Download the ZIP from this page. Unzip it into ~/.claude/skills/ to make the skill available to all your projects, or into a project's .claude/skills/ folder to limit it to that project.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with natural requests

You do not need to name the skill. The description lists the situations that should trigger it, so ask for the outcome in plain language. Mention Convex or a backend need and the agent should load the router.

If the project already has a Convex setup, say what you want to change. The router will send the request to the matching bundled skill or to the served catalog.

• "Start a new full-stack app with a Convex backend for a habit tracker."
• "Add Google sign-in to this Convex app."
• "Add Stripe subscriptions to my Convex project."
• "Review the functions under convex/ for security problems."
• "Add a nightly cron job that archives old rows."

Example workflows

For a new prototype, the agent should use convex-quickstart to scaffold a running app. For a feature on an existing app, it should use convex-add, which fetches current instructions for that capability from the served catalog. If the catalog cannot be reached, the source says the agent falls back to the bundled skill's own procedure.

The source also recommends running npx convex ai-files install first when a project has no Convex AI guidance or the guidance looks stale.

What you need before you start

The skill depends on the following. Check each item before asking for a routed task.

Billing capabilities need a Stripe account, but only if you use them.

• Node.js and npm, for npx convex commands
• A Convex project or account for the routed skills
• Internet access to the Convex capability catalog
• A Stripe account, only for billing capabilities

Safety notes: scripts and network

This skill does not run scripts of its own. It does use the network: convex-add fetches the served capability catalog at https://basic-anteater-667.convex.site/capabilities.json?src=agent-skills and each capability's doc from that host.

The source says served doc text is procedure instructions, not arbitrary shell to run blindly, so the agent should apply normal judgment. Capabilities marked tier>0, which spend money such as a domain purchase, always require your explicit confirmation.

Some routed skills change schemas, deploy previews or spend money. The source warns against using this on a live production app without review.

• Confirm any capability marked tier>0 before it proceeds.
• Review changes to schemas and deployments on a live app.
• Use convex-deploy-guard before deployment-affecting commands, per the bundled skill list.

Tips for better results

Tell the agent whether you are prototyping or running a live app, since the routed skills differ in risk. Ask for a review after the agent writes code, using convex-reviewer or convex-verify. If the agent seems to use outdated Convex syntax, ask it to check the current guidance or run the ai-files install command.

• State whether the app is a prototype or live.
• Ask for convex-verify after a feature is added.
• Keep production changes behind a review step.