Your agent doesn't need another integration. It needs a skill.

There are two ways to give an AI agent a new capability.

The first is an integration: a human finds the service, reads the docs, wires up a server, manages a key. The capability exists, but only because somebody plumbed it in first. That is how almost all agent tooling works today, and it is why most agents can do so little on their own.

The second way is a skill: a small instruction pack you install once. It sits in the agent's context and waits. The moment a task needs what it teaches, the agent already knows what to do, which endpoint to call, what to check before spending money, and what to hand back when it's done.

We just shipped PikaSim as the second kind:

npx skills add codebruinc/pikasim-mcp

That is the entire setup. Works with Claude Code, Codex, Cursor, Cline, and every other host that reads the open SKILL.md format. If you'd rather read it before installing, it's plain markdown on GitHub.

What your agent just learned

The cold start is the point

Most services assume someone signed up before the agent showed up. The skill assumes nothing. An agent that has never heard of PikaSim goes from zero to purchasing in four calls:

1. It calls one tool to mint its own wallet, and gets back a wallet code, an API key, and a small starting credit.
2. It saves both credentials. They are shown once, and they are the wallet.
3. It checks its balance. The starting credit often covers a first purchase, like an SMS code, before any funding at all.
4. It buys what the task needs and returns a working eSIM install link, a live phone number, or a verification code.

When the wallet runs low, the agent can create a deposit invoice and pay it over Lightning itself, or hand the payment page to its human. Either way, nobody fills out a form, because there is no form.

No MCP? Still works.

The skill's fallback path needs nothing but outbound HTTPS. The catalog is a public JSON API, purchases are authenticated REST calls, and even wallet creation works from a bare shell, because our MCP endpoint speaks plain JSON-RPC over POST. If your agent can run curl, it can put itself online.

Spending rules are baked in

A skill that moves real money should carry its own guardrails, so this one does. Installed as written, your agent will:

• Confirm the product and the exact live price with you before buying, unless you've explicitly told it to run autonomously.
• Never buy from a remembered price. It checks the catalog at decision time.
• Treat wallet codes and API keys as cash: never pasted into logs, commits, or shared channels.
• Deliver eSIMs as install links with QR codes, never as raw activation codes in chat.
• Know that a code from one specific service is never guaranteed on virtual numbers, and that quick verifications auto-refund if nothing arrives in 20 minutes, so it stops instead of burning money on retries.

And the wallet itself is the final guardrail: prepaid only, no credit line. Fund it with $5 and $5 is the worst case.

Why we built it this way

Agent capabilities are becoming something you install, not something you integrate. The skill format is open, portable across hosts, and readable by the person deciding whether to trust it. That is the right shape for a capability that spends money: inspectable instructions, a hard prepaid ceiling, and no credentials until the agent mints its own.

The agents don't need to prove they're human. They need to pay and go. Now yours knows how.