ngrok Agent Skills

A collection of skills for AI coding agents, covering how to expose local services, route and secure traffic, receive webhooks, and more.

Skills follow the Agent Skills format, so this repo works as a standalone skill pack and as a plugin for Claude Code, Cursor, ChatGPT, and Codex.

Installation

Any Agent Skills-compatible agent:

npx skills add ngrok/skills

Or install as a plugin:

• Claude Code (adds /ngrok:* slash commands):

  /plugin marketplace add ngrok/skills
  /plugin install ngrok@ngrok
  

• Cursor: Customize → Plugins → Add marketplace → From GitHub Repository → ngrok/skills.

• ChatGPT and Codex: run the command below, then restart and install ngrok from the Plugins Directory.

  codex plugin marketplace add ngrok/skills

Skills

Skill	Covers	Use when
expose-localhost	Give a local service a public URL — HTTP, or raw TCP for SSH, RDP, and databases	"Expose my local server", "Make my app publicly accessible", "SSH into a box with no public IP"
secure-endpoint	Auth, rate limiting, IP rules, or a custom response in front of an endpoint, without touching application code	"Require Google login on my ngrok URL", "Only allow my office IP range", "Put up a maintenance page"
receive-webhooks	Receive webhooks from Stripe, GitHub, and 70+ other providers, with signature verification at the edge	"Test Stripe webhooks locally", "One endpoint for webhooks from several providers"
run-mcp-gateway	Expose an MCP server under development so Claude, OpenAI, or another provider can connect to it, each with its own credential	"Let Claude connect to my local MCP server"
provision-sandbox-access	Give every sandbox, container, customer device, or tenant its own isolated endpoint, provisioned programmatically from a controlplane	"Each sandbox needs its own SSH access", "One endpoint per customer"
troubleshoot-ngrok	Diagnose any ngrok error from its docs page and your account's live state	"What does ERR_NGROK_3200 mean?", "Works on localhost, 400 through the ngrok URL"
ngrok-setup	Get authenticated and pointed at the right surface — CLI, SDK, API, or infrastructure-as-code	"Set up ngrok", "ngrok says I'm not authenticated"

Supporting References

Two skills are reference material rather than tasks. The skills above load them automatically — you do not invoke them directly, but they're installed as part of the set.

Skill	Covers
ngrok-engine	Endpoint types (agent, cloud, internal), HTTP vs TCP, and Traffic Policy — the rule language, plus a reference for all 26 traffic policy actions
ngrok-surfaces	How to apply configuration on each surface: agent CLI and ngrok.yml, the Go/JavaScript/Python/Rust SDKs, the REST API, Terraform, and the Kubernetes Operator

ngrok MCP server

When the ngrok MCP server is connected, the skills use it to check your account's live state, such as endpoints, agents, and traffic, instead of asking you. The plugin installs include the server. To sign in or set it up another way, see the ngrok MCP server docs.

Usage

Skills are automatically available once installed. The agent uses them when a relevant task is detected.

Examples:

Expose my app on port 3000 to the internet

Share my local server with Google OAuth so only people at @mycompany.com can access it

Receive Stripe webhooks and deliver them to a service that isn't publicly reachable

In Claude Code, you can also invoke a skill directly:

/ngrok:expose-localhost

Each skill contains:

• SKILL.md — instructions for the agent
• references/ — supporting documentation (optional)

Releases

Every merge to main that changes skills/ (or the bundle script) publishes a dated release, after the same checks as a pull request (.github/workflows/release.yml). A maintainer can also run the workflow by hand from main. The newest release is always at https://github.com/ngrok/skills/releases/latest.

Version format

• Tag. The tag is v followed by the version, for example v2026.11.04.
• Version. The version is the UTC date of the release run, written as YYYY.MM.DD with a zero-padded month and day.
  • The first release of a UTC day has no suffix.
  • Each later release that day appends .N, starting at .1: 2026.11.04, then 2026.11.04.1, then 2026.11.04.2.
• Ordering. Compare versions segment by segment as integers, and treat a missing segment as 0. So 2026.11.04 < 2026.11.04.1 < 2026.11.04.10 < 2026.11.05. Do not compare them as strings: .10 sorts before .9 as text.
• Not semver. The zero padding makes the version invalid semver. That is acceptable, because nothing reads these tags as semver. Do not copy the version into the plugin manifests' version fields. If a plugin format later needs a version, that is separate work.
• Strictly increasing. Every new tag must be greater than every existing v* tag. A same-day suffix and a UTC clock keep it that way. The workflow also checks it (scripts/next_version.py exits non-zero otherwise).

The reference implementation of the ordering rule is version_key() in scripts/next_version.py. Anything that compares skills versions should implement the same rule.

Assets

Each release carries two files:

Asset	Contents
skills.tar.gz	One top-level directory per skill, with that skill's files, and nothing else. There is no wrapping skills/ directory. Built with git archive of the committed skills/ tree by scripts/build_bundle.sh.
skills.tar.gz.sha256	One line in sha256sum format: <hex> skills.tar.gz.

Releases are immutable: once published, a release's assets and tag cannot change.

To fetch the newest release without the GitHub API, read the Location header of https://github.com/ngrok/skills/releases/latest without following it. It ends in /releases/tag/v<version>. Then download https://github.com/ngrok/skills/releases/download/v<version>/skills.tar.gz and skills.tar.gz.sha256.

License

MIT