mcpField
P
provision-sandbox-access

provision-sandbox-access

github.com/ngrokCloud & DevOps18141 installsUpdated 4d ago
Repository

About provision-sandbox-access

provision-sandbox-access is a skill for giving each sandbox, container, or customer appliance its own isolated ngrok entry point. It covers resource creation, scoped credentials, endpoint setup, and cleanup for platforms exposing untrusted workloads through a control plane.

Key features

  1. 01Create separate ngrok resources per workload
  2. 02Link resources and internal hostnames with a UUID
  3. 03Issue ACL-scoped agent credentials
  4. 04Forward to internal endpoints with Traffic Policy
  5. 05Delete resources and reconcile leftovers

Use cases

  1. 01Set up SSH access for ephemeral development sandboxes
  2. 02Isolate access for customer-premises devices
  3. 03Provision individual entry points for developer previews

MCP support

The ngrok MCP server can inspect resources created by the control plane, such as by listing endpoints. Provisioning itself is performed by the control plane through the ngrok API.

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/ngrok/skills --skill provision-sandbox-access

After installing, type/provision-sandbox-accessin Claude Code to confirm it loaded.

Bundled filesreferences/sandbox-ssh.md

Details

From the project README

Provision isolated access per sandbox

The job: a controlplane creates an environment - a sandbox, a container, a customer appliance - and each one needs its own reachable, isolated entry point. Not one shared endpoint with routing, but a separate set of ngrok resources per workload, created and destroyed with it.

This differs from expose-localhost in lifecycle and trust. There, a developer exposes their own machine and trusts themselves. Here, a platform exposes someone else's workload, at machine speed, and must assume the code inside it is hostile.

Before you start

You need an ngrok API key (not an agent authtoken) in the controlplane. See ngrok-setup. The ngrok MCP server does not replace it, because the controlplane is code that runs without you. If the MCP server is connected, use it to check what the controlplane created, such as gw_list_endpoints. The provisioning calls are documented in ngrok-surfaces (ngrok-surfaces/references/api.md and ngrok-surfaces/references/api-resources.md); this skill is the recipe, those are the mechanics.

The shape

public client
  -> cloud endpoint        (stable public URL/address, always on, runs the policy)
  -> forward-internal      (Traffic Policy)
  -> internal endpoint     (no public address; only reachable via forward-internal)
  -> ngrok agent inside the workload
  -> the local service (sshd, HTTP server, database)

The public side is provisioned by your controlplane before the workload boots. The internal side is created by the agent from inside the workload when it starts. They rendezvous on an agreed internal hostname, which is why that hostname has to be derived, not invented.

1. Derive a stable identifier

Pick the UUID your own object model already uses for whatever owns the environment - the sandbox record, the tenant, the session. Every ngrok resource you create is named and tagged from it.

Do not mint a new identifier for ngrok's benefit. The whole point is that ngrok's state can be reconciled against yours later, and that requires a shared key.

The internal hostname is derived from it too: tcp://sandbox-<uuid>.internal:22. Both sides of the rendezvous compute the same string from the same UUID, so neither has to discover the other.

2. Create the identity and address

In the controlplane, before the workload starts:

  • Service user (POST /service_users) - owns the credential, so it is not attached to an employee account. Note this resource takes only name, no metadata; put the identifier in the name.
  • A public address, which differs by protocol:
    • Raw TCP (SSH, RDP, a database) -> POST /reserved_addrs. ngrok assigns the host:port and returns it as addr. Takes a region (us unless the workload's users are elsewhere), fixed at creation.
    • HTTP/TLS -> POST /reserved_domains. You choose the hostname and it comes back as domain, not addr. No region - that field is deprecated for domains. A custom domain also needs its cname_target set in DNS before it resolves; an ngrok-owned subdomain needs no DNS step, which is usually what you want when provisioning per workload.

Fill description with something a human can identify in the dashboard and metadata with a JSON string carrying your UUID. See ngrok-surfaces for why both, and for the exact bodies.

3. Mint a credential that can do exactly one thing

This is the load-bearing security step.

{
  "description": "Agent authtoken for sandbox <uuid>",
  "metadata": "{\"sandbox_id\":\"<uuid>\"}",
  "acl": ["bind:tcp://sandbox-<uuid>.internal:22"],
  "owner_id": "<service user id>"
}

The bind: ACL restricts the token to creating that one endpoint and nothing else. Without an ACL the token can create any endpoint on your account - and this token is about to be placed inside a workload running untrusted code. Assume it will leak. With the ACL, a leaked token buys an attacker the ability to bind an endpoint that is already theirs.

The token comes back once and only once. Write it straight into the sandbox provider's secret store in the same code path; never log it or pass it back through your own API.

4. Create the cloud endpoint

POST /endpoints with type: "cloud", bindings: ["public"], and url: "tcp://{addr}" from step 2. Its Traffic Policy forwards to the internal endpoint the workload will later create:

on_tcp_connect:
  - actions:
      - type: forward-internal
        config:
          url: tcp://sandbox-<uuid>.internal:22

For HTTP workloads it is on_http_request and an https:// internal URL. The policy is passed to the API as a serialized string - see ngrok-surfaces.

This endpoint exists immediately and is stable. It will refuse connections until the agent inside the workload comes up and binds the internal endpoint, which is the correct behavior: the address is allocated at provision time, reachable at boot time.

Because the policy lives at the edge and not in the workload, this is also where you add controls the sandbox cannot remove. On a TCP endpoint you can add restrict-ips or deny in on_tcp_connect. On an HTTP endpoint, those actions are available, as well as oauth and rate-limit. See secure-endpoint for more details.

5. Configure the agent inside the workload

The workload needs: the authtoken in its environment, the ngrok agent binary, and a config file naming the internal endpoint.

version: 3
endpoints:
  - name: ssh
    # must match what the cloud endpoint forwards to, and what the ACL allows
    url: tcp://sandbox-<uuid>.internal:22
    upstream:
      url: 22
ngrok start --all --config /path/to/ngrok.yml

The authtoken comes from NGROK_AUTHTOKEN in the environment rather than the config file, so the token never lands on disk in the image. See ngrok-surfaces for the config surface generally, and references/sandbox-ssh.md for the full sandbox-container walkthrough including binary download and the embedded-SDK alternative.

Three strings must agree exactly: the internal URL in the agent config, the forward-internal target in the cloud endpoint's policy, and the bind: rule in the ACL. A mismatch fails as "the endpoint is up but nothing reaches it" or as an agent that cannot start at all. Generate all three from the same UUID in one function.

Tear it down

When the environment is destroyed, delete in reverse order: endpoint, credential, the reserved TCP address or HTTP domain, service user. Store each id against your record at create time.

Then run reconciliation anyway. Provisioning is a multi-call sequence that will fail partway through; a job that lists ngrok resources, parses metadata, and deletes those whose parent is gone is not optional at scale. ngrok-surfaces covers the delete calls.

Scenarios

Ephemeral dev sandboxes. The case above. SSH over a reserved TCP address, one per sandbox, torn down with the sandbox. Full walkthrough in references/sandbox-ssh.md.

Customer-premises devices or networks. Same resource set, longer-lived. One service user and ACL-scoped token per customer site so a compromise at one customer cannot touch another. Often several internal endpoints behind one agent (SSH, RDP, a database), each with its own cloud endpoint.

Per-developer preview environments. A wildcard reserved domain and one cloud endpoint can serve everyone, routing on hostname to <alias>.internal, with a per-developer token scoped bind:<alias>.internal. See secure-endpoint for the access-control side.

Notes for agents

  • Make the sequence idempotent and retry-safe, keyed on the caller's UUID. Partial provisioning is the normal failure mode.
  • Never suggest reusing one authtoken across tenants, and never suggest an unscoped token for a workload running user code. If the user proposes it, say plainly what it costs them.
  • Do not put the authtoken in a container image, a config file baked into the image, or logs. Environment variable from the provider's secret store.
  • Prefer the agent binary for a first implementation. Embedding ngrok in an existing process via an SDK (ngrok-surfaces) removes a moving part and may be the right end state, but it is harder to debug when the endpoint URL is wrong.
  • If the user is on Kubernetes, the operator may be a better fit than raw API calls; see ngrok-surfaces.

Comments

Markdown supported