The Applink Skill: AI coding agents that get telco integrations right

Ask a coding agent to add AppLink to your app and it will happily write if (response.ok) return "sent". AppLink answers failures with HTTP 200 and puts the real result in statusCode, so that line reports every error as a success. It will find /caas/direct/debit, get P1003 back and mark the order paid, when that call has only sent the subscriber an OTP. None of this is the model being careless. It is the model not having the contract.

What the skill is

The Applink Skill for Agents, built by hSenid Mobile, is that contract packaged for AI coding tools: every endpoint, every parameter and response field, all five callbacks, the published status codes, and the handful of rules that separate a working integration from a suspended one. It covers SMS, USSD, Subscription, OTP and CaaS charging, has no dependencies, and is derived field by field from the official API reference. Nothing in it is invented, and it records what the platform does not offer so an agent cannot fill the gap with a plausible guess.

It works in whatever language you already use. Every endpoint is written out as a runnable curl with each field defined, a language-neutral spec describes what surrounds the calls, and worked templates show the whole integration in TypeScript/Node, Python, Java, Go, PHP and C#.

Install it in your agent

In Claude Code it installs as a plugin, with its task skills available as slash commands:

/plugin marketplace add hSenidMobileCPaaS/Applink-as-a-Skill
/plugin install applink@applink

• Codex and the GitHub Copilot CLI install it from the same marketplace with their own plugin commands.
• Gemini CLI and Antigravity install it as an extension straight from the GitHub URL.
• Cursor, Windsurf, Cline, Kiro and Qoder take a single rules file copied into your project.
• Aider, Zed, Amp, Jules, Junie, OpenCode and anything else that reads AGENTS.md pick it up from the project root.
• No install at all: paste the raw AGENTS.md URL into any assistant and ask it to read the file.

The mistakes it prevents

• Deciding success on the HTTP status. The skill makes the agent read statusCode on every response.
• Treating P1003 as a completed charge. A CaaS charge is two calls, and the charging notification is the authoritative outcome.
• Retrying a failed verification by starting the charge again, which bills a real person twice. The fix is to re-prompt for the OTP against the same transaction.
• Sending destinationAddresses as a string. It is always an array.
• Hardcoding applicationId and password, calling AppLink from a browser or mobile app, or logging a callback body that contains your password.
• Keeping USSD sessions in an in-process map that works in development and breaks the moment you scale.

A contract the agent can query

Documentation still leaves an agent recalling field names from memory, so the whole contract also ships as structured data with a zero-dependency command-line tool over it. It gives the agent what an MCP server would, without running one: it can show any endpoint, print a runnable curl, explain a status code, diagnose a symptom, and validate a payload before it is ever sent. The tool makes no network calls and never sees a credential.

$ node tools/applink.mjs validate sms-send '{"message":"hi","destinationAddresses":"tel:8801959979376"}'

  ✗ 4 error(s)  against sms-send
    ✗ Missing required field "version"
    ✗ Missing required field "applicationId"
    ✗ Missing required field "password"
    ✗ "destinationAddresses" must be an ARRAY, got string.

$ node tools/applink.mjs diagnose "we charged the customer twice"
$ node tools/applink.mjs code P1003

Skills for each stage of the work

• applink-scaffold starts a new integration: config, credentials, client and typed models.
• applink-callbacks builds the inbound webhooks: SMS receive, delivery reports, USSD, subscription and charging notifications.
• applink-review audits existing code for the mistakes that cost money or leak credentials.
• applink-debug turns a status code or a symptom into a cause and a fix.
• applink-golive runs the pre-production checklist before you ask for approval.

From a prompt to production

Put your applicationId and password in an environment file and uncomment only the endpoints your app is provisioned for. Confirm your server's outbound IP is on the app's allow-list, run the smoke test, and exercise your callback handlers locally, with no AppLink account needed. A bundled mock of the platform checks every request body your code sends, whatever language sent it. Then ask your agent for the feature in plain words:

The skill is free to install and use, unmodified, for building and running your own AppLink integrations, and the code your agent writes with it is yours. As with any billing code, verify charging behaviour against the API reference before you go live.